واژه نامه یادگیری ماشینی

این واژه نامه، اصطلاحات کلی یادگیری ماشین، به اضافه اصطلاحات خاص TensorFlow را تعریف می کند.

آیا میدانستید؟

می‌توانید واژه‌نامه را با انتخاب موضوعی از فهرست بازشوی واژه‌نامه در نوار پیمایش بالا فیلتر کنید . نماد پرنده جوجه کشی تعاریفی را نشان می دهد که هدف آنها تازه واردان ML است.

آ

فرسایش

تکنیکی برای ارزیابی اهمیت یک ویژگی یا جزء با حذف موقت آن از یک مدل . سپس مدل را بدون آن ویژگی یا مؤلفه مجدداً آموزش می دهید، و اگر مدل بازآموزی شده به طور قابل توجهی بدتر عمل می کند، احتمالاً ویژگی یا مؤلفه حذف شده مهم است.

به عنوان مثال، فرض کنید یک مدل طبقه بندی را بر روی 10 ویژگی آموزش می دهید و به دقت 88 درصد در مجموعه آزمایشی دست می یابید. برای بررسی اهمیت ویژگی اول، می توانید مدل را تنها با استفاده از 9 ویژگی دیگر دوباره آموزش دهید. اگر مدل بازآموزی شده به طور قابل توجهی بدتر عمل کند (به عنوان مثال، دقت 55 درصد)، احتمالاً ویژگی حذف شده مهم است. برعکس، اگر مدل بازآموزی شده به همان اندازه خوب عمل کند، احتمالاً آن ویژگی چندان مهم نبوده است.

فرسایش همچنین می تواند به تعیین اهمیت موارد زیر کمک کند:

• اجزای بزرگتر، مانند کل زیرسیستم یک سیستم ML بزرگتر
• فرآیندها یا تکنیک ها، مانند مرحله پیش پردازش داده ها

در هر دو مورد، مشاهده خواهید کرد که چگونه عملکرد سیستم پس از حذف مؤلفه تغییر می کند (یا تغییر نمی کند).

تست A/B

روشی آماری برای مقایسه دو (یا بیشتر) تکنیک - A و B. به طور معمول، A یک تکنیک موجود است و B یک تکنیک جدید است. تست A/B نه تنها تعیین می‌کند که کدام تکنیک بهتر عمل می‌کند، بلکه مشخص می‌کند که آیا تفاوت از نظر آماری معنی‌دار است یا خیر.

تست A/B معمولاً یک معیار واحد را در دو تکنیک مقایسه می‌کند. به عنوان مثال، چگونه دقت مدل برای دو تکنیک مقایسه می شود؟ با این حال، آزمایش A/B می‌تواند هر تعداد محدودی از معیارها را با هم مقایسه کند.

تراشه شتاب دهنده

‎#GoogleCloud

دسته ای از اجزای سخت افزاری تخصصی که برای انجام محاسبات کلیدی مورد نیاز برای الگوریتم های یادگیری عمیق طراحی شده اند.

تراشه های شتاب دهنده (یا فقط شتاب دهنده ها ، به طور خلاصه) می توانند سرعت و کارایی وظایف آموزش و استنتاج را در مقایسه با یک CPU همه منظوره به میزان قابل توجهی افزایش دهند. آنها برای آموزش شبکه های عصبی و کارهای مشابه محاسباتی بسیار ایده آل هستند.

نمونه هایی از تراشه های شتاب دهنده عبارتند از:

• واحدهای پردازش تنسور Google ( TPU ) با سخت افزار اختصاصی برای یادگیری عمیق.
• پردازنده‌های گرافیکی NVIDIA که اگرچه در ابتدا برای پردازش گرافیکی طراحی شده‌اند، برای فعال کردن پردازش موازی طراحی شده‌اند که می‌تواند سرعت پردازش را به میزان قابل توجهی افزایش دهد.

دقت

#مبانی

تعداد پیش‌بینی‌های طبقه‌بندی صحیح تقسیم بر تعداد کل پیش‌بینی‌ها. به این معنا که:

$$\text{Accuracy} = \frac{\text{correct predictions}} {\text{correct predictions + incorrect predictions }}$$

به عنوان مثال، مدلی که 40 پیش‌بینی درست و 10 پیش‌بینی نادرست داشته باشد، دقتی برابر با:

$$\text{Accuracy} = \frac{\text{40}} {\text{40 + 10}} = \text{80%}$$

طبقه بندی باینری نام های خاصی را برای دسته های مختلف پیش بینی های صحیح و پیش بینی های نادرست ارائه می دهد. بنابراین، فرمول دقت برای طبقه بندی باینری به شرح زیر است:

$$\text{Accuracy} = \frac{\text{TP} + \text{TN}} {\text{TP} + \text{TN} + \text{FP} + \text{FN}}$$

جایی که:

• TP تعداد مثبت های واقعی (پیش بینی های صحیح) است.
• TN تعداد منفی های واقعی (پیش بینی های صحیح) است.
• FP تعداد مثبت کاذب (پیش‌بینی‌های نادرست) است.
• FN تعداد منفی های کاذب (پیش بینی های نادرست) است.

مقایسه و مقایسه دقت با دقت و یادآوری .

برای یادداشت های اضافی روی نماد کلیک کنید.

اگرچه برای برخی موقعیت‌ها یک معیار ارزشمند است، اما دقت برای برخی دیگر بسیار گمراه‌کننده است. قابل ذکر است که دقت معمولاً معیار ضعیفی برای ارزیابی مدل‌های طبقه‌بندی است که مجموعه داده‌های نامتعادل کلاس را پردازش می‌کنند.

برای مثال، فرض کنید در یک شهر نیمه گرمسیری خاص، تنها 25 روز در قرن برف می بارد. از آنجایی که روزهای بدون برف (طبقه منفی) بسیار بیشتر از روزهای با برف (طبقه مثبت) است، مجموعه داده های برف برای این شهر از نظر طبقه نامتعادل است. یک مدل طبقه‌بندی باینری را تصور کنید که قرار است هر روز برف یا بدون برف را پیش‌بینی کند، اما به سادگی هر روز «بدون برف» را پیش‌بینی می‌کند. این مدل بسیار دقیق است اما قدرت پیش بینی ندارد. جدول زیر نتایج یک قرن پیش‌بینی را خلاصه می‌کند:

دسته بندی	عدد
TP	0
TN	36500
FP	25
FN	0

بنابراین دقت این مدل عبارت است از:

accuracy = (TP + TN) / (TP + TN + FP + FN)
accuracy = (0 + 36500) / (0 + 36500 + 25 + 0) = 0.9993 = 99.93%

اگرچه دقت 99.93 درصد بسیار چشمگیر به نظر می رسد، این مدل در واقع قدرت پیش بینی ندارد.

دقت و یادآوری معمولاً معیارهای مفیدتری نسبت به دقت برای ارزیابی مدل‌های آموزش دیده بر روی مجموعه داده‌های نامتعادل کلاس هستند.

---

عمل

#rl

در یادگیری تقویتی ، مکانیزمی است که توسط آن عامل بین حالات محیط تغییر می کند. عامل با استفاده از یک خط مشی اقدام را انتخاب می کند.

عملکرد فعال سازی

#مبانی

تابعی که شبکه های عصبی را قادر می سازد تا روابط غیرخطی (پیچیده) بین ویژگی ها و برچسب را بیاموزند.

توابع فعال سازی محبوب عبارتند از:

• ReLU
• سیگموئید

نمودار توابع فعال سازی هرگز خطوط مستقیم منفرد نیستند. به عنوان مثال، نمودار تابع فعال سازی ReLU از دو خط مستقیم تشکیل شده است:

نمودار تابع فعال سازی سیگموئید به صورت زیر است:

برای مشاهده نمونه روی نماد کلیک کنید.

در یک شبکه عصبی، توابع فعال سازی مجموع وزنی تمام ورودی های یک نورون را دستکاری می کنند. برای محاسبه یک جمع وزنی، نورون حاصل جمع مقادیر و وزن های مربوطه را جمع می کند. برای مثال، فرض کنید ورودی مربوط به یک نورون شامل موارد زیر است:

مقدار ورودی	وزن ورودی
2	-1.3
-1	0.6
3	0.4

بنابراین مجموع وزنی این است:

weighted sum = (2)(-1.3) + (-1)(0.6) + (3)(0.4) = -2.0

فرض کنید طراح این شبکه عصبی تابع سیگموئید را به عنوان تابع فعال سازی انتخاب کند. در آن صورت، نورون سیگموئید -2.0 را محاسبه می کند که تقریباً 0.12 است. بنابراین، نورون 0.12 (به جای -2.0) را به لایه بعدی در شبکه عصبی منتقل می کند. شکل زیر بخش مربوط به فرآیند را نشان می دهد:

---

یادگیری فعال

یک رویکرد آموزشی که در آن الگوریتم برخی از داده هایی را که از آنها می آموزد انتخاب می کند . یادگیری فعال به ویژه زمانی ارزشمند است که نمونه‌های برچسب‌گذاری شده کمیاب یا گران هستند. به جای جستجوی کورکورانه طیف متنوعی از مثال‌های برچسب‌گذاری شده، یک الگوریتم یادگیری فعال به طور انتخابی به دنبال طیف خاصی از مثال‌هایی است که برای یادگیری نیاز دارد.

آداگراد

یک الگوریتم نزولی شیب پیچیده که گرادیان های هر پارامتر را مجدداً مقیاس می دهد و به طور مؤثر به هر پارامتر یک نرخ یادگیری مستقل می دهد. برای توضیح کامل، این مقاله را ببینید.

عامل

#rl

در یادگیری تقویتی ، نهادی که از سیاستی برای به حداکثر رساندن بازده مورد انتظار حاصل از انتقال بین حالت های محیط استفاده می کند.

خوشه بندی تجمعی

#خوشه بندی

به خوشه بندی سلسله مراتبی مراجعه کنید.

تشخیص ناهنجاری

فرآیند شناسایی نقاط پرت . به عنوان مثال، اگر میانگین برای یک ویژگی خاص 100 با انحراف استاندارد 10 باشد، تشخیص ناهنجاری باید مقدار 200 را به عنوان مشکوک علامت گذاری کند.

AR

مخفف واقعیت افزوده .

ناحیه زیر منحنی PR

به PR AUC (منطقه زیر منحنی PR) مراجعه کنید.

ناحیه زیر منحنی ROC

AUC (مساحت زیر منحنی ROC) را ببینید.

هوش عمومی مصنوعی

مکانیزمی غیر انسانی که طیف وسیعی از حل مسئله، خلاقیت و سازگاری را نشان می دهد. به عنوان مثال، برنامه ای که هوش عمومی مصنوعی را نشان می دهد می تواند متن را ترجمه کند، سمفونی بسازد و در بازی هایی که هنوز اختراع نشده اند برتری یابد.

هوش مصنوعی

#مبانی

یک برنامه یا مدل غیر انسانی که می تواند کارهای پیچیده را حل کند. برای مثال، برنامه یا مدلی که متن را ترجمه می کند یا برنامه یا مدلی که بیماری ها را از تصاویر رادیولوژیک شناسایی می کند، هر دو هوش مصنوعی را نشان می دهند.

به طور رسمی، یادگیری ماشین زیر شاخه هوش مصنوعی است. با این حال، در سال‌های اخیر، برخی از سازمان‌ها شروع به استفاده از اصطلاحات هوش مصنوعی و یادگیری ماشینی کرده‌اند.

توجه

#زبان

مکانیزمی که در شبکه عصبی استفاده می شود و اهمیت یک کلمه یا بخشی از یک کلمه را نشان می دهد. توجه، مقدار اطلاعاتی را که یک مدل برای پیش‌بینی رمز/کلمه بعدی نیاز دارد، فشرده می‌کند. یک مکانیسم توجه معمولی ممکن است شامل یک جمع وزنی بر روی مجموعه‌ای از ورودی‌ها باشد، جایی که وزن هر ورودی توسط بخش دیگری از شبکه عصبی محاسبه می‌شود.

رجوع به خود توجهی و خودتوجهی چند سر نیز شود که اجزای سازنده ترانسفورماتورها هستند.

صفت

#انصاف

مترادف برای ویژگی .

در عدالت یادگیری ماشینی، ویژگی ها اغلب به ویژگی های مربوط به افراد اشاره می کنند.

نمونه گیری ویژگی

#df

تاکتیکی برای آموزش یک جنگل تصمیم که در آن هر درخت تصمیم تنها یک زیرمجموعه تصادفی از ویژگی های ممکن را در هنگام یادگیری شرایط در نظر می گیرد. به طور کلی، زیر مجموعه متفاوتی از ویژگی ها برای هر گره نمونه برداری می شود. در مقابل، هنگام آموزش یک درخت تصمیم بدون نمونه گیری ویژگی، تمام ویژگی های ممکن برای هر گره در نظر گرفته می شود.

AUC (مساحت زیر منحنی ROC)

#مبانی

عددی بین 0.0 و 1.0 نشان دهنده توانایی یک مدل طبقه بندی باینری برای جداسازی کلاس های مثبت از کلاس های منفی است. هر چه AUC به 1.0 نزدیکتر باشد، مدل توانایی بهتری برای جداسازی کلاس ها از یکدیگر دارد.

برای مثال، تصویر زیر یک مدل طبقه‌بندی کننده را نشان می‌دهد که کلاس‌های مثبت (بیضی سبز) را از کلاس‌های منفی (مستطیل‌های بنفش) کاملاً جدا می‌کند. این مدل غیرواقعی کامل دارای AUC 1.0 است:

برعکس، تصویر زیر نتایج یک مدل طبقه‌بندی کننده را نشان می‌دهد که نتایج تصادفی ایجاد می‌کند. این مدل دارای AUC 0.5 است:

بله، مدل قبلی دارای AUC 0.5 است، نه 0.0.

اکثر مدل ها جایی بین دو حالت افراطی هستند. به عنوان مثال، مدل زیر موارد مثبت را تا حدودی از منفی جدا می کند و بنابراین دارای AUC بین 0.5 و 1.0 است:

AUC هر مقداری را که برای آستانه طبقه بندی تنظیم کرده اید نادیده می گیرد. در عوض، AUC تمام آستانه های طبقه بندی ممکن را در نظر می گیرد.

برای اطلاع از رابطه بین منحنی های AUC و ROC روی نماد کلیک کنید.

AUC نشان دهنده سطح زیر منحنی ROC است. به عنوان مثال، منحنی ROC برای مدلی که به طور کامل نکات مثبت را از منفی جدا می کند، به صورت زیر است:

AUC ناحیه خاکستری در تصویر قبل است. در این حالت غیر معمول، مساحت به سادگی طول ناحیه خاکستری (1.0) ضرب در عرض ناحیه خاکستری (1.0) است. بنابراین، حاصل ضرب 1.0 و 1.0 AUC دقیقاً 1.0 را به دست می دهد که بالاترین امتیاز AUC ممکن است.

برعکس، منحنی ROC برای طبقه‌بندی‌کننده‌ای که اصلاً نمی‌تواند کلاس‌ها را از هم جدا کند، به شرح زیر است. مساحت این منطقه خاکستری 0.5 است.

یک منحنی معمولی ROC تقریباً شبیه زیر است:

محاسبه مساحت زیر این منحنی به صورت دستی دشوار خواهد بود، به همین دلیل است که یک برنامه معمولاً بیشتر مقادیر AUC را محاسبه می کند.

---

برای تعریف رسمی تر AUC روی نماد کلیک کنید.

AUC احتمال این است که یک طبقه بندی کننده مطمئن تر از مثبت بودن یک مثال مثبت تصادفی انتخاب شده باشد تا اینکه یک مثال منفی تصادفی انتخاب شده مثبت باشد.

---

واقعیت افزوده

#تصویر

فناوری ای که یک تصویر تولید شده توسط کامپیوتر را بر روی دید کاربر از دنیای واقعی قرار می دهد و بنابراین یک نمای ترکیبی ارائه می دهد.

رمزگذار خودکار

#زبان

#تصویر

سیستمی که یاد می گیرد مهم ترین اطلاعات را از ورودی استخراج کند. رمزگذارهای خودکار ترکیبی از رمزگذار و رمزگشا هستند. رمزگذارهای خودکار بر فرآیند دو مرحله ای زیر متکی هستند:

1. رمزگذار ورودی را به یک قالب (معمولا) با ابعاد پایین تر (متوسط) ترسیم می کند.
2. رمزگشا با نگاشت قالب با ابعاد پایین تر به فرمت ورودی با ابعاد بالاتر، یک نسخه با اتلاف از ورودی اصلی ایجاد می کند.

رمزگذارهای خودکار با تلاش رمزگشا برای بازسازی ورودی اصلی از فرمت میانی رمزگذار تا حد امکان به صورت سرتاسر آموزش داده می شوند. از آنجایی که فرمت میانی کوچکتر (بُعد پایین) از فرمت اصلی است، رمزگذار خودکار مجبور است اطلاعاتی را که در ورودی ضروری است، بیاموزد و خروجی کاملاً با ورودی یکسان نخواهد بود.

مثلا:

• اگر داده های ورودی گرافیکی باشد، کپی غیر دقیق شبیه به گرافیک اصلی است، اما تا حدودی تغییر یافته است. شاید کپی غیر دقیق، نویز را از گرافیک اصلی حذف می کند یا برخی از پیکسل های از دست رفته را پر می کند.
• اگر داده ورودی متن باشد، یک رمزگذار خودکار متن جدیدی تولید می کند که متن اصلی را تقلید می کند (اما مشابه نیست).

رمزگذارهای خودکار متغیر را نیز ببینید.

تعصب اتوماسیون

#انصاف

هنگامی که یک تصمیم گیرنده انسانی توصیه های ارائه شده توسط یک سیستم تصمیم گیری خودکار را بر اطلاعاتی که بدون اتوماسیون انجام می شود ترجیح می دهد، حتی زمانی که سیستم تصمیم گیری خودکار خطا می کند.

AutoML

هر فرآیند خودکار برای ساخت مدل های یادگیری ماشینی . AutoML می تواند به طور خودکار کارهایی مانند موارد زیر را انجام دهد:

• مناسب ترین مدل را جستجو کنید.
• تنظیم هایپرپارامترها
• داده ها را آماده کنید (از جمله انجام مهندسی ویژگی ).
• مدل به دست آمده را مستقر کنید.

AutoML برای دانشمندان داده مفید است زیرا می تواند در زمان و تلاش آنها در توسعه خطوط لوله یادگیری ماشین صرفه جویی کند و دقت پیش بینی را بهبود بخشد. همچنین برای افراد غیرمتخصص مفید است، زیرا وظایف پیچیده یادگیری ماشینی را برای آنها در دسترس تر می کند.

کتابخانه‌های

مدل خود رگرسیون

#زبان

#تصویر

#تولید هوش مصنوعی

مدلی که پیش بینی را بر اساس پیش بینی های قبلی خود استنباط می کند. برای مثال، مدل‌های زبان رگرسیون خودکار، نشانه‌های بعدی را بر اساس نشانه‌های پیش‌بینی‌شده قبلی پیش‌بینی می‌کنند. همه مدل‌های زبان بزرگ مبتنی بر Transformer دارای رگرسیون خودکار هستند.

در مقابل، مدل‌های تصویر مبتنی بر GAN معمولاً رگرسیون خودکار نیستند، زیرا آنها یک تصویر را در یک گذر به جلو و نه به صورت تکراری در مراحل تولید می‌کنند. با این حال، برخی از مدل‌های تولید تصویر دارای رگرسیون خودکار هستند ، زیرا آنها یک تصویر را در مراحل تولید می‌کنند.

از دست دادن کمکی

یک تابع از دست دادن - که در ارتباط با تابع اصلی از دست دادن مدل شبکه عصبی استفاده می شود - که به تسریع تمرین در طی تکرارهای اولیه هنگامی که وزن ها به طور تصادفی مقداردهی اولیه می شوند، کمک می کند.

توابع از دست دادن کمکی ، گرادیان های موثر را به لایه های قبلی هدایت می کنند. این امر با مبارزه با مشکل شیب ناپدید شدن، همگرایی را در طول تمرین تسهیل می کند.

دقت متوسط

معیاری برای خلاصه کردن عملکرد یک دنباله رتبه بندی شده از نتایج. دقت متوسط ​​با در نظر گرفتن میانگین مقادیر دقت برای هر نتیجه مرتبط (هر نتیجه در فهرست رتبه‌بندی شده که در آن فراخوان نسبت به نتیجه قبلی افزایش می‌یابد) محاسبه می‌شود.

همچنین به ناحیه زیر منحنی PR مراجعه کنید.

وضعیت هم تراز با محور

#df

در درخت تصمیم ، شرایطی که فقط شامل یک ویژگی است. برای مثال، اگر ناحیه یک ویژگی است، شرایط زیر یک شرط تراز محور است:

area > 200

کنتراست با حالت مایل .

ب

پس انتشار

#مبانی

الگوریتمی که نزول گرادیان را در شبکه های عصبی پیاده سازی می کند.

آموزش یک شبکه عصبی شامل تکرارهای زیادی از چرخه دو پاس زیر است:

1. در طول پاس رو به جلو ، سیستم دسته‌ای از نمونه‌ها را پردازش می‌کند تا پیش‌بینی (های) را به دست آورد. سیستم هر پیش بینی را با هر برچسب مقایسه می کند. تفاوت بین مقدار پیش‌بینی و برچسب، ضرر آن مثال است. سیستم تلفات را برای همه نمونه‌ها جمع‌آوری می‌کند تا مجموع ضرر را برای دسته فعلی محاسبه کند.
2. در طول گذر به عقب (انتشار عقب)، سیستم با تنظیم وزن تمام نورون ها در تمام لایه(های) پنهان، تلفات را کاهش می دهد.

شبکه‌های عصبی اغلب حاوی نورون‌های زیادی در لایه‌های پنهان بسیاری هستند. هر یک از این نورون ها به روش های مختلفی در از دست دادن کلی نقش دارند. انتشار معکوس تعیین می کند که آیا وزن اعمال شده روی نورون های خاص افزایش یا کاهش یابد.

نرخ یادگیری یک ضریب است که میزان افزایش یا کاهش هر وزنه توسط هر پاس به عقب را کنترل می کند. نرخ یادگیری زیاد هر وزن را بیش از یک نرخ یادگیری کوچک افزایش یا کاهش می دهد.

در اصطلاح حساب دیفرانسیل و انتگرال، پس انتشار قانون زنجیره حساب دیفرانسیل و انتگرال را پیاده سازی می کند. یعنی پس انتشار مشتق جزئی خطا را با توجه به هر پارامتر محاسبه می کند. برای جزئیات بیشتر، این آموزش را در Machine Learning Crash Course ببینید.

سال‌ها پیش، تمرین‌کنندگان ML مجبور بودند کدی را برای پیاده‌سازی انتشار پس‌انداز بنویسند. API های مدرن ML مانند TensorFlow اکنون پس انتشار را برای شما پیاده سازی می کنند. اوه!

بسته بندی

#df

روشی برای آموزش یک گروه که در آن هر مدل سازنده بر روی یک زیرمجموعه تصادفی از نمونه‌های آموزشی نمونه‌برداری شده با جایگزینی تمرین می‌کند. به عنوان مثال، یک جنگل تصادفی مجموعه ای از درختان تصمیم گیری است که با کیسه بندی آموزش دیده اند.

اصطلاح bagging مخفف b ootstrap agg regat ing است.

کیسه کلمات

#زبان

نمایشی از کلمات در یک عبارت یا متن، صرف نظر از ترتیب. به عنوان مثال، کیسه کلمات سه عبارت زیر را به طور یکسان نشان می دهد:

• سگ می پرد
• سگ را می پرد
• سگ می پرد

هر کلمه به یک شاخص در یک بردار پراکنده نگاشت می شود، که در آن بردار برای هر کلمه در واژگان یک شاخص دارد. به عنوان مثال، عبارت dog jumps در یک بردار ویژگی با مقادیر غیر صفر در سه شاخص مربوط به کلمات the , dog و jumps نگاشت می شود. مقدار غیر صفر می تواند یکی از موارد زیر باشد:

• A 1 برای نشان دادن وجود یک کلمه.
• تعداد دفعاتی که یک کلمه در کیسه ظاهر می شود. به عنوان مثال، اگر عبارت were the maroon dog یک سگ با خز قهوه ای است ، هر دو قهوه ای و سگ به صورت 2 نشان داده می شوند، در حالی که کلمات دیگر به عنوان 1 نمایش داده می شوند.
• مقدار دیگری مانند لگاریتم تعداد دفعاتی که یک کلمه در کیسه ظاهر می شود.

خط پایه

مدلی که به عنوان یک نقطه مرجع برای مقایسه عملکرد یک مدل دیگر (معمولاً پیچیده تر) استفاده می شود. به عنوان مثال، یک مدل رگرسیون لجستیک ممکن است به عنوان یک پایه خوب برای یک مدل عمیق عمل کند.

برای یک مشکل خاص، خط مبنا به توسعه دهندگان مدل کمک می کند تا حداقل عملکرد مورد انتظاری را که یک مدل جدید باید به آن دست پیدا کند تا مدل جدید مفید باشد، کمّی کنند.

دسته ای

#مبانی

مجموعه مثال های مورد استفاده در یک تکرار آموزشی. اندازه دسته تعداد نمونه ها را در یک دسته تعیین می کند.

برای توضیح نحوه ارتباط یک دسته با یک دوره، به epoch مراجعه کنید.

استنتاج دسته ای

#TensorFlow

‎#GoogleCloud

فرآیند استنباط پیش‌بینی‌ها در چندین مثال بدون برچسب که به زیرمجموعه‌های کوچک‌تر ("مجموعه") تقسیم می‌شوند.

استنتاج دسته ای می تواند از ویژگی های موازی سازی تراشه های شتاب دهنده استفاده کند. به این معنا که شتاب‌دهنده‌های متعدد می‌توانند به طور همزمان پیش‌بینی‌هایی را بر روی دسته‌های مختلف نمونه‌های بدون برچسب استنباط کنند و تعداد استنتاج‌ها را به‌طور چشمگیری در هر ثانیه افزایش دهند.

عادی سازی دسته ای

عادی کردن ورودی یا خروجی توابع فعالسازی در یک لایه پنهان . نرمال سازی دسته ای می تواند مزایای زیر را ارائه دهد:

• با محافظت در برابر وزن های دورتر ، شبکه های عصبی را پایدارتر کنید.
• نرخ های یادگیری بالاتر را فعال کنید، که می تواند آموزش را سرعت بخشد.
• بیش از حد برازش را کاهش دهید.

اندازه دسته

#مبانی

تعداد نمونه ها در یک دسته . به عنوان مثال، اگر اندازه دسته 100 باشد، مدل در هر تکرار 100 نمونه را پردازش می کند.

استراتژی های اندازه دسته ای محبوب زیر هستند:

• نزول گرادیان تصادفی (SGD) که در آن اندازه دسته 1 است.
• دسته کامل، که در آن اندازه دسته تعداد نمونه ها در کل مجموعه آموزشی است. به عنوان مثال، اگر مجموعه آموزشی حاوی یک میلیون مثال باشد، اندازه دسته ای یک میلیون نمونه خواهد بود. دسته کامل معمولا یک استراتژی ناکارآمد است.
• مینی بچ که در آن اندازه دسته معمولا بین 10 تا 1000 است. مینی بچ معمولا کارآمدترین استراتژی است.

شبکه عصبی بیزی

یک شبکه عصبی احتمالی که عدم قطعیت در وزن ها و خروجی ها را محاسبه می کند. یک مدل رگرسیون شبکه عصبی استاندارد معمولاً یک مقدار اسکالر را پیش‌بینی می‌کند . به عنوان مثال، یک مدل استاندارد قیمت خانه را 853000 پیش بینی می کند. در مقابل، یک شبکه عصبی بیزی توزیع مقادیر را پیش بینی می کند. به عنوان مثال، یک مدل بیزی قیمت خانه 853000 را با انحراف معیار 67200 پیش بینی می کند.

یک شبکه عصبی بیزی برای محاسبه عدم قطعیت در وزن ها و پیش بینی ها بر قضیه بیز متکی است. یک شبکه عصبی بیزی زمانی می‌تواند مفید باشد که تعیین کمیت عدم قطعیت، مانند مدل‌های مرتبط با داروها، مهم باشد. شبکه‌های عصبی بیزی نیز می‌توانند به جلوگیری از برازش بیش از حد کمک کنند.

بهینه سازی بیزی

یک تکنیک مدل رگرسیون احتمالی برای بهینه‌سازی توابع هدف محاسباتی گران‌قیمت با بهینه‌سازی جایگزینی که عدم قطعیت را از طریق تکنیک یادگیری بیزی تعیین می‌کند. از آنجایی که بهینه سازی بیزی به خودی خود بسیار گران است، معمولاً برای بهینه سازی وظایف گران قیمت که دارای تعداد کمی پارامتر هستند، مانند انتخاب فراپارامترها استفاده می شود.

معادله بلمن

#rl

در یادگیری تقویتی، هویت زیر با تابع Q بهینه برآورده می شود:

\[Q(s, a) = r(s, a) + \gamma \mathbb{E}_{s'|s,a} \max_{a'} Q(s', a')\]

الگوریتم های یادگیری تقویتی این هویت را برای ایجاد یادگیری Q از طریق قانون به روز رسانی زیر اعمال می کنند:

\[Q(s,a) \gets Q(s,a) + \alpha \left[r(s,a) + \gamma \displaystyle\max_{\substack{a_1}} Q(s’,a’) - Q(s,a) \right] \]

فراتر از یادگیری تقویتی، معادله بلمن کاربردهایی در برنامه نویسی پویا دارد. به مدخل ویکی پدیا برای معادله بلمن مراجعه کنید.

BERT (نمایش رمزگذار دوطرفه از ترانسفورماتورها)

#زبان

معماری مدلی برای نمایش متن یک مدل BERT آموزش دیده می تواند به عنوان بخشی از یک مدل بزرگتر برای طبقه بندی متن یا سایر وظایف ML عمل کند.

BERT دارای ویژگی های زیر است:

• از معماری ترانسفورماتور استفاده می کند و بنابراین به خود توجهی متکی است.
• از بخش رمزگذار ترانسفورماتور استفاده می کند. وظیفه رمزگذار به جای انجام یک کار خاص مانند طبقه بندی، تولید بازنمایی متن خوب است.
• دو جهته است.
• از ماسک برای تمرینات بدون نظارت استفاده می کند.

انواع BERT عبارتند از:

• ALBERT که مخفف A L ight BERT است.
• LaBSE .

برای مروری بر BERT به منبع باز BERT: پیش‌آموزش پیشرفته برای پردازش زبان طبیعی مراجعه کنید.

تعصب (اخلاق / انصاف)

#انصاف

#مبانی

1. کلیشه سازی، تعصب یا طرفداری نسبت به برخی چیزها، افراد یا گروه ها نسبت به دیگران. این سوگیری ها می توانند بر جمع آوری و تفسیر داده ها، طراحی یک سیستم و نحوه تعامل کاربران با یک سیستم تأثیر بگذارند. اشکال این نوع سوگیری عبارتند از:

• تعصب اتوماسیون
• سوگیری تایید
• تعصب آزمایشگر
• سوگیری اسناد گروهی
• سوگیری ضمنی
• تعصب درون گروهی
• سوگیری همگنی برون گروهی

2. خطای سیستماتیک معرفی شده توسط یک روش نمونه گیری یا گزارش. اشکال این نوع سوگیری عبارتند از:

• سوگیری پوشش
• سوگیری بدون پاسخ
• سوگیری مشارکت
• سوگیری گزارش
• سوگیری نمونه گیری
• سوگیری انتخاب

نباید با اصطلاح سوگیری در مدل‌های یادگیری ماشین یا سوگیری پیش‌بینی اشتباه گرفته شود.

تعصب (ریاضی) یا اصطلاح سوگیری

#مبانی

رهگیری یا جبران از مبدأ. تعصب یک پارامتر در مدل های یادگیری ماشینی است که با یکی از موارد زیر نشان داده می شود:

• ب
• w ₀

به عنوان مثال، بایاس b در فرمول زیر است:

$$y' = b + w_1x_1 + w_2x_2 + … w_nx_n$$

در یک خط دوبعدی ساده، بایاس فقط به معنای «قطعه y» است. به عنوان مثال، بایاس خط در تصویر زیر 2 است.

تعصب وجود دارد زیرا همه مدل ها از مبدا (0,0) شروع نمی شوند. به عنوان مثال، فرض کنید یک پارک تفریحی برای ورود به آن 2 یورو و برای هر ساعت اقامت مشتری 0.5 یورو اضافی هزینه دارد. بنابراین، مدلی که هزینه کل را نگاشت می کند، بایاس 2 دارد زیرا کمترین هزینه 2 یورو است.

سوگیری نباید با سوگیری در اخلاق و انصاف یا سوگیری پیش بینی اشتباه شود.

دو طرفه

#زبان

اصطلاحی که برای توصیف سیستمی استفاده می‌شود که متنی را که هم قبل و هم بعد از یک بخش هدف از متن است، ارزیابی می‌کند. در مقابل، یک سیستم یک طرفه فقط متنی را که قبل از بخش هدفی از متن قرار دارد ارزیابی می کند.

به عنوان مثال، یک مدل زبان پوشانده شده را در نظر بگیرید که باید احتمالات کلمه یا کلماتی را که نشان دهنده خط زیر در سؤال زیر هستند تعیین کند:

_____ با شما چیست؟

یک مدل زبانی یک طرفه باید احتمالات خود را فقط بر اساس زمینه ارائه شده توسط کلمات "What"، "is" و "the" استوار کند. در مقابل، یک مدل زبان دوطرفه همچنین می‌تواند زمینه را از «با» و «شما» به دست آورد، که ممکن است به مدل کمک کند پیش‌بینی‌های بهتری ایجاد کند.

مدل زبان دو طرفه

#زبان

یک مدل زبان که احتمال وجود یک نشانه داده شده در یک مکان معین در گزیده ای از متن را بر اساس متن قبلی و بعدی تعیین می کند.

بیگرام

#دنباله

#زبان

یک N گرم که در آن N=2 است.

طبقه بندی باینری

#مبانی

یک نوع کار طبقه بندی که یکی از دو کلاس منحصر به فرد را پیش بینی می کند:

• طبقه مثبت
• طبقه منفی

به عنوان مثال، دو مدل یادگیری ماشین زیر، هر کدام دسته بندی باینری را انجام می دهند:

• مدلی که تعیین می‌کند پیام‌های ایمیل هرزنامه هستند (کلاس مثبت) یا اسپم نیستند (کلاس منفی).
• مدلی که علائم پزشکی را ارزیابی می کند تا مشخص کند آیا یک فرد دارای یک بیماری خاص (طبقه مثبت) است یا آن بیماری (طبقه منفی) را ندارد.

در تقابل با طبقه بندی چند طبقه .

همچنین به رگرسیون لجستیک و آستانه طبقه بندی مراجعه کنید.

شرایط باینری

#df

در درخت تصمیم ، شرایطی که فقط دو نتیجه ممکن دارد، معمولاً بله یا خیر . به عنوان مثال، شرایط زیر یک شرط باینری است:

temperature >= 100

کنتراست با شرط غیر باینری .

بنینگ

مترادف کلمه سطل .

BLEU (دو زبانه ارزیابی ارزشیابی)

#زبان

امتیازی بین 0.0 تا 1.0، شامل کیفیت ترجمه بین دو زبان انسانی (مثلاً بین انگلیسی و روسی) است. نمره BLEU 1.0 نشان دهنده ترجمه کامل است. نمره BLEU 0.0 نشان دهنده یک ترجمه وحشتناک است.

افزایش

یک تکنیک یادگیری ماشینی که به طور مکرر مجموعه‌ای از طبقه‌بندی‌کننده‌های ساده و نه چندان دقیق (که به آنها طبقه‌بندی‌کننده‌های ضعیف گفته می‌شود) را در یک طبقه‌بندی با دقت بالا (یک طبقه‌بندی‌کننده قوی) با بالا بردن نمونه‌هایی که مدل در حال حاضر طبقه‌بندی اشتباه می‌کند، ترکیب می‌کند.

جعبه مرزی

#تصویر

در یک تصویر، مختصات ( x ، y ) یک مستطیل در اطراف ناحیه مورد نظر، مانند سگ در تصویر زیر.

صدا و سیما

گسترش شکل یک عملوند در یک عملیات ریاضی ماتریس به ابعاد سازگار با آن عملیات. برای مثال، جبر خطی مستلزم آن است که دو عملوند در یک عملیات جمع ماتریس باید ابعاد یکسانی داشته باشند. در نتیجه، نمی توانید یک ماتریس شکل (m, n) را به بردار با طول n اضافه کنید. پخش این عملیات را با گسترش عملی بردار طول n به یک ماتریس شکل (m, n) با تکرار همان مقادیر در پایین هر ستون امکان پذیر می کند.

به عنوان مثال، با توجه به تعاریف زیر، جبر خطی A+B را ممنوع می کند زیرا A و B دارای ابعاد مختلف هستند:

A = [[7, 10, 4],
     [13, 5, 9]]
B = [2]

با این حال، پخش عملیات A+B را با گسترش عملی B به موارد زیر فعال می کند:

 [[2, 2, 2],
  [2, 2, 2]]

بنابراین، A+B اکنون یک عملیات معتبر است:

[[7, 10, 4],  +  [[2, 2, 2],  =  [[ 9, 12, 6],
 [13, 5, 9]]      [2, 2, 2]]      [15, 7, 11]]

برای جزئیات بیشتر به شرح زیر در مورد پخش در NumPy مراجعه کنید.

سطل سازی

#مبانی

تبدیل یک ویژگی واحد به چندین ویژگی باینری به نام سطل یا bins ، که معمولاً بر اساس یک محدوده مقدار است. ویژگی خرد شده معمولاً یک ویژگی پیوسته است.

به عنوان مثال، به جای نمایش دما به عنوان یک ویژگی ممیز شناور منفرد، می توانید محدوده دما را به سطل های مجزا تقسیم کنید، مانند:

• <= 10 درجه سانتیگراد سطل "سرد" خواهد بود.
• 11 تا 24 درجه سانتیگراد سطل "معتدل" خواهد بود.
• >= 25 درجه سانتیگراد سطل "گرم" خواهد بود.

مدل با هر مقدار در یک سطل یکسان رفتار می کند. به عنوان مثال، مقادیر 13 و 22 هر دو در سطل معتدل هستند، بنابراین مدل با دو مقدار یکسان رفتار می کند.

برای یادداشت های اضافی روی نماد کلیک کنید.

اگر دما را به عنوان یک ویژگی پیوسته نشان دهید، مدل دما را به عنوان یک ویژگی واحد در نظر می گیرد. اگر دما را به عنوان سه سطل نشان دهید، مدل هر سطل را به عنوان یک ویژگی جداگانه در نظر می گیرد. یعنی یک مدل می تواند روابط جداگانه هر سطل را با برچسب یاد بگیرد. به عنوان مثال، یک مدل رگرسیون خطی می تواند وزن های جداگانه ای را برای هر سطل یاد بگیرد.

افزایش تعداد سطل ها با افزایش تعداد روابطی که مدل شما باید یاد بگیرد، مدل شما را پیچیده تر می کند. به عنوان مثال، سطل های سرد، معتدل و گرم اساساً سه ویژگی مجزا برای مدل شما هستند که می توانند با آن تمرین کنند. اگر تصمیم دارید دو سطل دیگر اضافه کنید - به عنوان مثال، انجماد و داغ - مدل شما اکنون باید روی پنج ویژگی جداگانه آموزش ببیند.

چگونه می دانید که چند سطل ایجاد کنید یا محدوده هر سطل چقدر باید باشد؟ پاسخ‌ها معمولاً به مقدار کافی آزمایش نیاز دارند.

---

سی

لایه کالیبراسیون

یک تعدیل پس از پیش‌بینی، معمولاً برای در نظر گرفتن سوگیری پیش‌بینی . پیش‌بینی‌ها و احتمالات تنظیم‌شده باید با توزیع مجموعه‌ای از برچسب‌ها مطابقت داشته باشند.

نسل نامزد

#سیستم های بازگردانی

مجموعه اولیه توصیه هایی که توسط یک سیستم توصیه انتخاب شده است. به عنوان مثال، کتابفروشی را در نظر بگیرید که 100000 عنوان کتاب عرضه می کند. مرحله تولید نامزد، فهرست بسیار کوچکتری از کتاب‌های مناسب برای یک کاربر خاص، مثلاً 500، ایجاد می‌کند. اما حتی 500 کتاب برای توصیه به کاربر بسیار زیاد است. مراحل بعدی و گرانتر یک سیستم توصیه (مانند امتیاز دهی و رتبه بندی مجدد ) این 500 را به مجموعه ای بسیار کوچکتر و مفیدتر از توصیه ها کاهش می دهد.

نمونه گیری نامزد

بهینه سازی زمان آموزش که احتمال آن را برای همه برچسب های مثبت محاسبه می کند ، به عنوان مثال SoftMax ، اما فقط برای یک نمونه تصادفی از برچسب های منفی. به عنوان مثال ، با توجه به نمونه ای با برچسب Beagle و Dog ، نمونه گیری نامزد احتمالات پیش بینی شده و شرایط ضرر مربوطه را برای: محاسبه می کند:

• بیگل
• سگ
• یک زیر مجموعه تصادفی از کلاسهای منفی باقیمانده (به عنوان مثال ، CAT ، Lollipop ، حصار ).

ایده این است که طبقات منفی می توانند از تقویت منفی کمتر مکرر یاد بگیرند تا زمانی که طبقات مثبت همیشه تقویت مثبت مثبت شوند ، و این در واقع به صورت تجربی مشاهده می شود.

نمونه گیری کاندیداها از نظر محاسباتی نسبت به الگوریتم های آموزش که پیش بینی را برای همه کلاس های منفی محاسبه می کنند ، به ویژه هنگامی که تعداد کلاس های منفی بسیار زیاد باشد ، محاسباتی است.

داده های طبقه بندی

#فونداستال ها

ویژگی های دارای مجموعه خاصی از مقادیر ممکن. به عنوان مثال ، یک ویژگی طبقه بندی شده به نام traffic-light-state را در نظر بگیرید ، که فقط می تواند یکی از سه مقدار ممکن زیر را داشته باشد:

• red
• yellow
• green

با نمایندگی از traffic-light-state به عنوان یک ویژگی طبقه بندی ، یک مدل می تواند تأثیرات متفاوت از رفتار red ، green و yellow را بر رفتار راننده بیاموزد.

ویژگی های طبقه بندی شده گاهی اوقات ویژگی های گسسته ای نامیده می شوند.

تضاد با داده های عددی .

الگوی زبان علی

#زبان

مترادف برای مدل زبان یک طرفه .

برای تضاد رویکردهای مختلف جهت در مدل سازی زبان ، به مدل زبان دو طرفه مراجعه کنید.

نقطه مرکزی

#خوشه ای

مرکز یک خوشه که توسط یک الگوریتم k-means یا k-median تعیین می شود. به عنوان مثال ، اگر k 3 باشد ، الگوریتم K- یا k-dedian 3 Centroids پیدا می کند.

خوشه بندی مبتنی بر سانتروئید

#خوشه ای

دسته ای از الگوریتم های خوشه بندی که داده ها را در خوشه های غیرشی سازماندهی می کنند. K-Means پرکاربردترین الگوریتم خوشه بندی مبتنی بر سانتروئید است.

تضاد با الگوریتم های خوشه بندی سلسله مراتبی .

تکیه گاه به فکر

#زبان

#Generativeai

یک تکنیک سریع مهندسی که یک مدل بزرگ زبان (LLM) را برای توضیح استدلال خود ، گام به گام تشویق می کند. به عنوان مثال ، فوری زیر را در نظر بگیرید ، با توجه ویژه به جمله دوم:

یک راننده در یک اتومبیل که از 0 تا 60 مایل در ساعت در 7 ثانیه می رود ، چند نیرو G را تجربه می کند؟ در پاسخ ، تمام محاسبات مربوطه را نشان دهید.

پاسخ LLM احتمالاً:

• دنباله ای از فرمول های فیزیک را نشان دهید ، و مقادیر 0 ، 60 و 7 را در مکان های مناسب وصل کنید.
• توضیح دهید که چرا این فرمول ها را انتخاب کرده و متغیرهای مختلف را به چه معنی دارد.

با فکر زنجیره ای ، LLM را وادار می کند تا تمام محاسبات را انجام دهد ، که ممکن است منجر به پاسخ صحیح تر شود. علاوه بر این ، برانگیختگی زنجیره ای از کاربر ، کاربر را قادر می سازد تا مراحل LLM را بررسی کند تا مشخص شود که آیا این پاسخ معنی دارد یا خیر.

ایست بازرسی

داده هایی که وضعیت پارامترهای یک مدل را در یک تکرار آموزش خاص ضبط می کند. پاسگاه ها وزن مدل صادرات یا انجام آموزش در چندین جلسه را فعال می کنند. پاسگاه ها همچنین آموزش را برای ادامه خطاهای گذشته (به عنوان مثال ، پیشگیری شغلی) قادر می سازد.

هنگام تنظیم دقیق ، نقطه شروع برای آموزش مدل جدید یک بازرسی خاص از مدل از پیش آموزش دیده خواهد بود.

کلاس

#فونداستال ها

مقوله ای که یک برچسب می تواند به آن تعلق داشته باشد. مثلا:

• در یک مدل طبقه بندی باینری که اسپم را تشخیص می دهد ، ممکن است دو کلاس هرزنامه باشد و نه اسپم .
• در یک مدل طبقه بندی چند کلاس که نژادهای سگ را مشخص می کند ، کلاس ها ممکن است پودل ، بیگل ، میناکاری و غیره باشند.

یک مدل طبقه بندی یک کلاس را پیش بینی می کند. در مقابل ، یک مدل رگرسیون به جای یک کلاس یک عدد را پیش بینی می کند.

مدل طبقه بندی

#فونداستال ها

مدلی که پیش بینی آن یک کلاس است. به عنوان مثال ، موارد زیر همه مدل های طبقه بندی هستند:

• مدلی که زبان یک جمله ورودی را پیش بینی می کند (فرانسوی؟ اسپانیایی؟ ایتالیایی؟).
• مدلی که گونه های درختی را پیش بینی می کند (افرا؟ بلوط؟ باباب؟).
• مدلی که کلاس مثبت یا منفی را برای یک وضعیت پزشکی خاص پیش بینی می کند.

در مقابل ، مدل های رگرسیون تعداد را به جای کلاس ها پیش بینی می کنند.

دو نوع متداول از مدل های طبقه بندی عبارتند از:

• طبقه بندی دوتایی
• طبقه بندی چند طبقه

آستانه طبقه بندی

#فونداستال ها

در یک طبقه بندی باینری ، تعدادی بین 0 تا 1 که خروجی خام یک مدل رگرسیون لجستیک را به یک پیش بینی کلاس مثبت یا کلاس منفی تبدیل می کند. توجه داشته باشید که آستانه طبقه بندی ارزشی است که یک انسان انتخاب می کند ، نه یک ارزش انتخاب شده توسط آموزش مدل.

یک مدل رگرسیون لجستیک یک مقدار خام را بین 0 و 1 به دست می آورد.

• اگر این مقدار خام از آستانه طبقه بندی بیشتر باشد ، کلاس مثبت پیش بینی می شود.
• اگر این مقدار خام کمتر از آستانه طبقه بندی باشد ، کلاس منفی پیش بینی می شود.

به عنوان مثال ، فرض کنید آستانه طبقه بندی 0.8 است. اگر مقدار خام 0.9 باشد ، مدل کلاس مثبت را پیش بینی می کند. اگر مقدار خام 0.7 باشد ، مدل کلاس منفی را پیش بینی می کند.

انتخاب آستانه طبقه بندی به شدت بر تعداد مثبت کاذب و منفی های دروغین تأثیر می گذارد.

برای یادداشت های اضافی روی نماد کلیک کنید.

با تکامل مدل ها یا مجموعه داده ها ، مهندسان گاهی اوقات آستانه طبقه بندی را نیز تغییر می دهند. هنگامی که آستانه طبقه بندی تغییر می کند ، پیش بینی های کلاس مثبت به طور ناگهانی می توانند کلاس های منفی و برعکس شوند.

به عنوان مثال ، یک مدل پیش بینی بیماری طبقه بندی باینری را در نظر بگیرید. فرض کنید وقتی سیستم در سال اول اجرا می شود:

• مقدار خام برای یک بیمار خاص 0.95 است.
• آستانه طبقه بندی 0.94 است.

بنابراین ، سیستم کلاس مثبت را تشخیص می دهد. (بیمار گاز می گیرد ، "اوه نه! من بیمار هستم!")

یک سال بعد ، شاید ارزشها اکنون به شرح زیر باشد:

• مقدار خام برای همان بیمار در 0.95 باقی مانده است.
• آستانه طبقه بندی به 0.97 تغییر می کند.

بنابراین ، سیستم اکنون آن بیمار را به عنوان کلاس منفی مجدداً معرفی می کند. ("روز مبارک! من بیمار نیستم.") همان بیمار. تشخیص متفاوت.

---

مجموعه داده های کلاس متعادل

#فونداستال ها

یک مجموعه داده برای یک مشکل طبقه بندی که در آن تعداد کل برچسب های هر کلاس به طور قابل توجهی متفاوت است. به عنوان مثال ، یک مجموعه داده طبقه بندی باینری را در نظر بگیرید که دو برچسب آن به شرح زیر تقسیم شده است:

• 1000،000 برچسب منفی
• 10 برچسب مثبت

نسبت برچسب های منفی به مثبت 100000 به 1 است ، بنابراین این یک مجموعه داده با کلاس تعادل است.

در مقابل ، مجموعه داده های زیر از کلاس تعادل نیست زیرا نسبت برچسب های منفی به برچسب های مثبت نسبتاً نزدیک به 1 است:

• 517 برچسب منفی
• 483 برچسب مثبت

مجموعه داده های چند طبقه نیز می توانند کلاس تعادل باشند. به عنوان مثال ، مجموعه داده های طبقه بندی چند کلاس زیر نیز با کلاس تعادل است زیرا یک برچسب نمونه های بسیار بیشتری نسبت به دو مورد دیگر دارد:

• 1،000،000 برچسب با کلاس "سبز"
• 200 برچسب با کلاس "بنفش"
• 350 برچسب با کلاس "نارنجی"

همچنین به آنتروپی ، کلاس اکثریت و کلاس اقلیت مراجعه کنید.

بریدن

#فونداستال ها

تکنیکی برای دستیابی به خارج از کشور با انجام یا هر دو مورد زیر:

• کاهش مقادیر ویژگی که بیشتر از حداکثر آستانه در آن آستانه حداکثر است.
• افزایش مقادیر ویژگی که کمتر از حداقل آستانه تا حداقل آستانه است.

به عنوان مثال ، فرض کنید که <0.5 ٪ مقادیر برای یک ویژگی خاص در خارج از محدوده 40-60 قرار دارد. در این حالت ، می توانید موارد زیر را انجام دهید:

• کلیپ تمام مقادیر بیش از 60 (حداکثر آستانه) دقیقاً 60 باشد.
• کلیپ تمام مقادیر زیر 40 (حداقل آستانه) دقیقاً 40 باشد.

Outliers می تواند به مدل ها آسیب برساند و گاهی اوقات باعث افزایش وزن در هنگام تمرین می شود. برخی از فضای باز نیز می توانند معیارهایی مانند دقت را خراب کنند. قطع یک روش متداول برای محدود کردن آسیب است.

مقادیر شیب شیب شیب در طول آموزش در محدوده مشخص شده است.

Cloud TPU

#tensorflow

#GOOGLECLOUD

یک شتاب دهنده سخت افزاری تخصصی که برای سرعت بخشیدن به بارهای یادگیری ماشین در Google Cloud Platform طراحی شده است.

خوشه بندی

#خوشه ای

نمونه های مرتبط با گروه بندی ، به ویژه در هنگام یادگیری بدون نظارت . پس از گروه بندی همه نمونه ها ، یک انسان می تواند به صورت اختیاری به معنای هر خوشه باشد.

بسیاری از الگوریتم های خوشه بندی وجود دارند. به عنوان مثال ، خوشه های الگوریتم K-Means بر اساس نزدیکی آنها به یک مرکز سرود ، مانند نمودار زیر:

سپس یک محقق انسانی می تواند خوشه ها را مرور کند و به عنوان مثال ، خوشه 1 را به عنوان "درختان کوتوله" و خوشه 2 به عنوان "درختان با اندازه کامل" بررسی کند.

به عنوان مثال دیگر ، یک الگوریتم خوشه بندی را بر اساس فاصله یک مثال از یک نقطه مرکزی در نظر بگیرید ، به شرح زیر است:

همکاری مشترک

هنگامی که نورون ها به جای تکیه بر رفتار شبکه به عنوان یک کل ، با تکیه بر خروجی های نورون های خاص دیگر ، الگوهای آموزش داده ها را پیش بینی می کنند. هنگامی که الگوهای ایجاد همبستگی در داده های اعتبار سنجی وجود ندارند ، پس سازگاری باعث ایجاد بیش از حد می شود. تنظیم مجدد ترک تحصیل باعث کاهش همبستگی می شود زیرا ترک تحصیل تضمین می کند که سلولهای عصبی نمی توانند صرفاً به سلولهای عصبی خاص دیگر متکی باشند.

فیلتر مشارکتی

#سیستم ها

پیش بینی در مورد علایق یک کاربر بر اساس علایق بسیاری از کاربران دیگر. از فیلتر مشارکتی اغلب در سیستم های توصیه استفاده می شود.

وضعیت

#DF

در یک درخت تصمیم ، هر گره ای که یک عبارت را ارزیابی می کند. به عنوان مثال ، بخش زیر از یک درخت تصمیم گیری شامل دو شرط است:

یک شرط نیز تقسیم یا آزمایش نامیده می شود.

شرایط کنتراست با برگ .

همچنین ببینید:

• وضعیت دودویی
• وضعیت غیر دودویی
• شرط محور
• تهوع

پیکربندی

فرایند اختصاص مقادیر اولیه خاصیت مورد استفاده برای آموزش یک مدل ، از جمله:

• لایه های آهنگساز مدل
• محل داده ها
• HyperParameters مانند:
  • میزان یادگیری
  • تکرارها
  • بهینه ساز
  • تابع ضرر

در پروژه های یادگیری ماشین ، پیکربندی را می توان از طریق یک فایل پیکربندی ویژه یا از طریق کتابخانه های پیکربندی مانند موارد زیر انجام داد:

• هرجلندی
• جین
• کمانچه

سوگیری تایید

#فریبی

تمایل به جستجوی ، تفسیر ، لطف و به یاد آوردن اطلاعات به روشی که اعتقادات یا فرضیه های موجود را تأیید می کند. توسعه دهندگان یادگیری ماشین ممکن است سهواً داده ها را به روش هایی که بر نتیجه حمایت از اعتقادات موجود خود تأثیر می گذارد ، جمع آوری یا برچسب گذاری کند. تعصب تأیید نوعی تعصب ضمنی است.

تعصب آزمایشگر نوعی تعصب تأیید است که در آن یک آزمایشگر مدلهای آموزش را ادامه می دهد تا اینکه یک فرضیه موجود تأیید شود.

ماتریس سردرگمی

#فونداستال ها

یک جدول NXN که تعداد پیش بینی های صحیح و نادرست را که یک مدل طبقه بندی ساخته شده است خلاصه می کند. به عنوان مثال ، ماتریس سردرگمی زیر را برای یک مدل طبقه بندی باینری در نظر بگیرید:

	تومور (پیش بینی شده)	غیر تومور (پیش بینی شده)
تومور (حقیقت زمین)	18 (TP)	1 (FN)
غیر تومور (حقیقت زمین)	6 (FP)	452 (TN)

ماتریس سردرگمی قبلی موارد زیر را نشان می دهد:

• از 19 پیش بینی که حقیقت زمین تومور بود ، مدل به درستی 18 طبقه بندی شده و به طور نادرست طبقه بندی می شود 1.
• از 458 پیش بینی که در آن حقیقت زمین غیر تومور بود ، مدل به درستی 452 طبقه بندی و نادرست طبقه بندی می شود.

ماتریس سردرگمی برای یک مشکل طبقه بندی چند طبقه می تواند به شما در شناسایی الگوهای اشتباه کمک کند. به عنوان مثال ، ماتریس سردرگمی زیر را برای یک مدل طبقه بندی چند طبقه 3 طبقه که سه نوع عنبیه مختلف را طبقه بندی می کند (Virginica ، Versicolor و Setosa) در نظر بگیرید. هنگامی که حقیقت زمین ویرجینیکا بود ، ماتریس سردرگمی نشان می دهد که این مدل به احتمال زیاد به اشتباه پیش بینی Versicolor از Setosa را پیش بینی می کند:

	setosa (پیش بینی شده)	بررسیکولور (پیش بینی شده)	Virginica (پیش بینی شده)
setosa (حقیقت زمین)	88	12	0
برقی (حقیقت زمین)	6	141	7
ویرجینیکا (حقیقت زمین)	2	27	109

به عنوان مثال دیگر ، یک ماتریس سردرگمی می تواند نشان دهد که مدلی که برای تشخیص رقم های دستنویس آموزش دیده است ، تمایل دارد به جای 4 به اشتباه 9 را پیش بینی کند ، یا به اشتباه 1 را به جای 7 پیش بینی کند.

ماتریس سردرگمی حاوی اطلاعات کافی برای محاسبه انواع معیارهای عملکرد از جمله دقت و فراخوان است.

حوزه انتخابیه

#زبان

تقسیم یک جمله به ساختارهای گرامری کوچکتر ("ترکیبات"). قسمت بعدی سیستم ML ، مانند یک مدل درک زبان طبیعی ، می تواند ترکیبات را راحت تر از جمله اصلی تجزیه کند. برای مثال جمله زیر را در نظر بگیرید:

دوست من دو گربه را پذیرفت.

یک تجزیه کننده حوزه انتخابیه می تواند این جمله را به دو مؤلفه زیر تقسیم کند:

• دوست من یک عبارت اسم است.
• دو گربه پذیرفته شده یک عبارت فعل است.

این ترکیبات را می توان بیشتر به ترکیبات کوچکتر تقسیم کرد. به عنوان مثال ، عبارت فعل

دو گربه را پذیرفت

می توان بیشتر به زیر تقسیم شد:

• اتخاذ شده یک فعل است.
• دو گربه یک عبارت دیگر است.

ویژگی مداوم

#فونداستال ها

یک ویژگی نقطه شناور با دامنه نامتناهی از مقادیر ممکن ، مانند دما یا وزن.

تضاد با ویژگی گسسته .

نمونه گیری آسان

به منظور انجام آزمایشات سریع ، استفاده از مجموعه داده های علمی جمع نشده است. بعداً ، تغییر به یک مجموعه داده جمع آوری شده علمی ضروری است.

همگرایی

#فونداستال ها

هنگامی که مقادیر ضرر با هر تکرار بسیار کم تغییر کند یا اصلاً تغییر کند. به عنوان مثال ، منحنی ضرر زیر همگرایی را در حدود 700 تکرار نشان می دهد:

هنگامی که آموزش های اضافی مدل را بهبود نمی بخشد ، یک مدل همگرا می شود .

در یادگیری عمیق ، ارزش های از دست دادن گاهی اوقات برای بسیاری از تکرارها قبل از نزول ، ثابت یا تقریباً تقریباً باقی می مانند. در یک دوره طولانی از مقادیر ضرر مداوم ، ممکن است به طور موقت حس کاذب از همگرایی بدست آورید.

همچنین توقف زود هنگام را ببینید.

تابع محدب

تابعی که در آن منطقه بالاتر از نمودار عملکرد یک مجموعه محدب است. عملکرد محدب نمونه اولیه چیزی مانند حرف U شکل گرفته است. به عنوان مثال ، موارد زیر همه توابع محدب هستند:

در مقابل ، عملکرد زیر محدب نیست. توجه کنید که چگونه منطقه بالای نمودار یک مجموعه محدب نیست:

یک عملکرد کاملاً محدب دقیقاً یک نقطه حداقل محلی دارد ، که همچنین حداقل نقطه جهانی است. توابع کلاسیک U شکل عملکردهای کاملاً محدب هستند. با این حال ، برخی از توابع محدب (به عنوان مثال ، خطوط مستقیم) U شکل نیستند.

برای نگاهی عمیق تر به ریاضی روی نماد کلیک کنید.

بسیاری از عملکردهای از دست دادن رایج ، از جمله موارد زیر ، توابع محدب هستند:

• L ₂ ضرر
• از دست دادن ورود
• l ₁ منظم سازی
• تنظیم منظم L ₂

بسیاری از تغییرات نزول شیب تضمین شده است تا نقطه ای نزدیک به حداقل عملکرد کاملاً محدب پیدا کند. به طور مشابه ، بسیاری از تغییرات نزول شیب تصادفی احتمال بالایی (هرچند ضمانت نه) برای یافتن نقطه نزدیک به حداقل عملکرد دقیق محدب است.

مجموع دو عملکرد محدب (به عنوان مثال ، تنظیم مجدد L ₂ Loss + L ₁ ) یک تابع محدب است.

مدل های عمیق هرگز توابع محدب نیستند. نکته قابل توجه ، الگوریتم هایی که برای بهینه سازی محدب طراحی شده اند ، به هر حال تمایل به یافتن راه حل های منطقی خوب در شبکه های عمیق دارند ، حتی اگر این راه حل ها تضمین نمی شوند که حداقل جهانی باشند.

---

بهینه سازی محدب

فرایند استفاده از تکنیک های ریاضی مانند نزول شیب برای یافتن حداقل عملکرد محدب . تحقیقات زیادی در زمینه یادگیری ماشینی بر تدوین مشکلات مختلف به عنوان مشکلات بهینه سازی محدب و در حل کارآمدتر این مشکلات متمرکز شده است.

برای جزئیات کامل ، به بهینه سازی Boyd و Vandenberghe ، بهینه سازی محدب مراجعه کنید.

مجموعه محدب

زیر مجموعه ای از فضای اقلیدسی به گونه ای که خطی بین هر دو نقطه در زیر مجموعه کاملاً در زیر مجموعه باقی می ماند. به عنوان مثال ، دو شکل زیر مجموعه محدب هستند:

در مقابل ، دو شکل زیر مجموعه محدب نیستند:

پیچیدگی

#تصویر

در ریاضیات ، به طور اتفاقی ، ترکیبی از دو کارکرد. در یادگیری ماشین ، یک حلقوی فیلتر حلقوی و ماتریس ورودی را به منظور آموزش وزنه ها مخلوط می کند.

اصطلاح "Convolution" در یادگیری ماشین اغلب یک روش کوتاه برای مراجعه به عملکرد حلقوی یا لایه حلقوی است.

بدون پیچیدگی ، یک الگوریتم یادگیری ماشین باید برای هر سلول در یک تانسور بزرگ ، یک وزن جداگانه یاد بگیرد. به عنوان مثال ، یک آموزش الگوریتم یادگیری ماشین بر روی تصاویر 2K x 2K مجبور به یافتن وزن 4M جداگانه می شود. به لطف پیچیدگی ها ، یک الگوریتم یادگیری ماشین فقط باید برای هر سلول در فیلتر حلقوی ، وزنه هایی پیدا کند و حافظه مورد نیاز برای آموزش مدل را به طرز چشمگیری کاهش می دهد. هنگامی که فیلتر حلقوی اعمال می شود ، به سادگی در سلولها تکرار می شود به گونه ای که هر یک توسط فیلتر ضرب می شود.

فیلتر حلقوی

#تصویر

یکی از دو بازیگر در یک عملیات حلقوی . (بازیگر دیگر یک قطعه از یک ماتریس ورودی است.) یک فیلتر حلقوی یک ماتریس است که دارای رتبه مشابه ماتریس ورودی است ، اما شکل کمتری دارد. به عنوان مثال ، با توجه به یک ماتریس ورودی 28x28 ، فیلتر می تواند هر ماتریس 2D کوچکتر از 28x28 باشد.

در دستکاری عکاسی ، تمام سلولهای موجود در یک فیلتر حلقوی به طور معمول به یک الگوی ثابت از آنهایی و صفرها تنظیم می شوند. در یادگیری ماشین ، فیلترهای حلقوی به طور معمول با شماره های تصادفی بذر می شوند و سپس شبکه مقادیر ایده آل را آموزش می دهد .

لایه کانولوشن

#تصویر

لایه ای از یک شبکه عصبی عمیق که در آن یک فیلتر حلقوی در امتداد یک ماتریس ورودی عبور می کند. به عنوان مثال ، فیلتر Convolutional 3x3 زیر را در نظر بگیرید:

انیمیشن زیر یک لایه حلقوی متشکل از 9 عملیات حلقوی شامل ماتریس ورودی 5x5 را نشان می دهد. توجه کنید که هر عملیات حلقوی روی یک قطعه 3x3 متفاوت از ماتریس ورودی کار می کند. ماتریس 3x3 حاصل (در سمت راست) شامل نتایج 9 عملیات حلقوی است:

شبکه عصبی کانولوشنال

#تصویر

یک شبکه عصبی که در آن حداقل یک لایه یک لایه حلقوی است. یک شبکه عصبی Convolutional معمولی شامل ترکیبی از لایه های زیر است:

• لایه های حلقوی
• لایه های استخر
• لایه های متراکم

شبکه های عصبی Convolutional در انواع خاصی از مشکلات مانند تشخیص تصویر موفقیت بزرگی داشته اند.

عمل جغرافیایی

#تصویر

عملیات ریاضی دو مرحله ای زیر:

1. ضرب عناصر عاقلانه فیلتر حلقوی و یک قطعه از ماتریس ورودی. (برش ماتریس ورودی دارای رتبه و اندازه یکسان با فیلتر حلقوی است.)
2. جمع بندی تمام مقادیر موجود در ماتریس محصول حاصل.

به عنوان مثال ، ماتریس ورودی 5x5 زیر را در نظر بگیرید:

اکنون فیلتر Convolutional 2x2 زیر را تصور کنید:

هر عملیات حلقوی شامل یک قطعه 2x2 منفرد از ماتریس ورودی است. به عنوان مثال ، فرض کنید ما از برش 2x2 در سمت چپ ماتریس ورودی استفاده می کنیم. بنابراین ، عملیات Convolution در این برش به شرح زیر است:

یک لایه حلقوی از یک سری عملیات حلقوی تشکیل شده است که هر یک بر روی یک قطعه متفاوت از ماتریس ورودی عمل می کنند.

هزینه

مترادف برای از دست دادن .

آموزش مشترک

یک رویکرد یادگیری نیمه تحت نظارت به ویژه هنگامی که همه شرایط زیر صحیح باشد مفید است:

• نسبت نمونه های بدون برچسب به نمونه های دارای برچسب در مجموعه داده زیاد است.
• این یک مشکل طبقه بندی ( باینری یا چند طبقه ) است.
• مجموعه داده شامل دو مجموعه مختلف از ویژگی های پیش بینی کننده است که از یکدیگر مستقل و مکمل هستند.

آموزش مشترک اساساً سیگنال های مستقل را به یک سیگنال قوی تر تقویت می کند. به عنوان مثال ، یک مدل طبقه بندی را در نظر بگیرید که اتومبیل های استفاده شده فردی را خوب یا بد طبقه بندی می کند. یک مجموعه از ویژگی های پیش بینی کننده ممکن است بر خصوصیات کل مانند سال ، ساخت و مدل ماشین متمرکز شود. مجموعه دیگری از ویژگی های پیش بینی کننده ممکن است روی سابقه رانندگی مالک قبلی و سابقه نگهداری خودرو متمرکز شود.

مقاله اصلی در مورد آموزش همزمان با ترکیب داده های برچسب زده و بدون برچسب با آموزش همزمان توسط بلوم و میچل ترکیب شده است.

انصاف ضد خلاف

#فریبی

یک متریک انصاف که بررسی می کند که آیا طبقه بندی کننده نتیجه یکسان را برای یک فرد تولید می کند ، همانطور که برای فرد دیگری که با اولی یکسان است ، به جز با توجه به یک یا چند ویژگی حساس است. ارزیابی یک طبقه بندی کننده برای انصاف ضد عملی ، یکی از روشها برای گسترش منابع بالقوه تعصب در یک مدل است.

به "هنگامی که جهان ها با هم برخورد می کنند: ادغام فرضیات ضد خلاف مختلف در انصاف" برای بحث بیشتر در مورد انصاف ضد خلاف.

تعصب پوشش

#فریبی

به تعصب انتخاب مراجعه کنید.

شکوفه تصادف

#زبان

یک جمله یا عبارت با معنای مبهم. شکوفه های تصادف مشکل مهمی در درک زبان طبیعی دارند. به عنوان مثال ، نوار قرمز تیتر آسمان خراش یک شکوفه خرابی است زیرا یک مدل NLU می تواند تیتر را به معنای واقعی کلمه یا تصویری تفسیر کند.

برای یادداشت های اضافی روی نماد کلیک کنید.

فقط برای روشن کردن آن تیتر مرموز:

• نوار قرمز می تواند به هر یک از موارد زیر مراجعه کند:
  • چسبنده
  • بوروکراسی بیش از حد
• نگه داشتن می تواند به هر یک از موارد زیر مراجعه کند:
  • پشتیبانی ساختاری
  • تاخیر

---

منتقد

#RL

مترادف برای شبکه Q Deep .

فاش کردن

تعمیم از دست دادن ورود به مشکلات طبقه بندی چند طبقه . آنتروپی متقاطع تفاوت بین دو توزیع احتمال را تعیین می کند. همچنین به گیج کننده مراجعه کنید.

اعتبار سنجی متقابل

مکانیسمی برای تخمین اینکه چگونه یک مدل با آزمایش مدل در برابر یک یا چند زیر مجموعه داده غیر همپوشانی از مجموعه آموزش ، به داده های جدید تعمیم می یابد.

D

تحلیل داده ها

با در نظر گرفتن نمونه ها ، اندازه گیری و تجسم ، درک داده ها را بدست آورید. تجزیه و تحلیل داده ها می تواند به ویژه در هنگام دریافت یک مجموعه داده ، قبل از اینکه مدل اول را بسازد ، مفید باشد. همچنین در درک آزمایشات و مشکلات اشکال زدایی با سیستم بسیار مهم است.

افزایش داده ها

#تصویر

با تبدیل نمونه های موجود برای ایجاد نمونه های اضافی ، دامنه و تعداد نمونه های آموزشی را افزایش می دهد. به عنوان مثال ، فرض کنید تصاویر یکی از ویژگی های شما هستند ، اما مجموعه داده شما حاوی نمونه های تصویر کافی برای مدل برای یادگیری انجمن های مفید نیست. در حالت ایده آل ، شما می توانید تصاویر برچسب زده شده به اندازه کافی را به مجموعه داده خود اضافه کنید تا مدل خود بتواند به درستی آموزش ببیند. اگر این امکان پذیر نباشد ، افزایش داده ها می توانند هر تصویر را بچرخانند ، کشش دهند و منعکس شوند تا انواع بسیاری از تصویر اصلی را تولید کنند ، احتمالاً داده های دارای برچسب به اندازه کافی را به دست می آورند تا بتوانند آموزش عالی را فعال کنند.

قاب داده

#فونداستال ها

یک داده Pandas محبوب برای نمایش مجموعه داده ها در حافظه.

DataFrame مشابه یک جدول یا صفحه گسترده است. هر ستون از یک DataFrame دارای یک نام (یک هدر) است و هر سطر توسط یک عدد منحصر به فرد مشخص می شود.

هر ستون در یک DataFrame مانند یک آرایه 2D ساختار یافته است ، به جز اینکه به هر ستون می توان نوع داده خود را اختصاص داد.

همچنین به صفحه مرجع رسمی pandas.dataframe مراجعه کنید.

موازی سازی داده ها

روشی برای مقیاس بندی آموزش یا استنتاج که یک مدل کامل را بر روی چندین دستگاه تکرار می کند و سپس زیر مجموعه ای از داده های ورودی را به هر دستگاه منتقل می کند. موازی سازی داده ها می توانند آموزش و استنباط را در اندازه های دسته ای بسیار بزرگ فعال کنند. با این حال ، موازی سازی داده ها مستلزم آن است که مدل به اندازه کافی کوچک باشد تا در همه دستگاه ها جای بگیرد.

موازی سازی داده ها به طور معمول آموزش و استنباط را سرعت می بخشد.

همچنین به موازی سازی مدل مراجعه کنید.

مجموعه داده یا مجموعه داده

#فونداستال ها

مجموعه ای از داده های خام ، معمولاً (اما نه منحصراً) در یکی از قالب های زیر سازماندهی شده است:

• یک صفحه گسترده
• پرونده ای در قالب CSV (مقادیر جدا از کاما)

API DataSet (tf.data)

#tensorflow

یک API Tensorflow با سطح بالا برای خواندن داده ها و تبدیل آن به شکلی که یک الگوریتم یادگیری ماشین به آن نیاز دارد. یک شی tf.data.Dataset دنباله ای از عناصر را نشان می دهد ، که در آن هر عنصر حاوی یک یا چند تنش است. یک شیء tf.data.Iterator دسترسی به عناصر یک Dataset را فراهم می کند.

برای جزئیات بیشتر در مورد API DataSet ، به tf.data مراجعه کنید: خطوط لوله ورودی tensorflow را در راهنمای برنامه نویس TensorFlow بسازید.

مرز تصمیم گیری

جداکننده بین کلاسهای آموخته شده توسط یک مدل در یک کلاس باینری یا مشکلات طبقه بندی چند طبقه . به عنوان مثال ، در تصویر زیر که یک مشکل طبقه بندی باینری را نشان می دهد ، مرز تصمیم گیری مرز بین کلاس نارنجی و کلاس آبی است:

جنگل تصمیم

#DF

مدلی که از درختان تصمیم گیری چندگانه ایجاد شده است. یک جنگل تصمیم گیری با تجمع پیش بینی های درختان تصمیم خود پیش بینی می کند. انواع محبوب جنگل های تصمیم گیری شامل جنگل های تصادفی و درختان تقویت شده شیب است .

آستانه تصمیم گیری

مترادف برای آستانه طبقه بندی .

درخت تصمیم

#DF

یک مدل یادگیری تحت نظارت متشکل از مجموعه ای از شرایط و ترک سلسله مراتب سازمان یافته. به عنوان مثال ، موارد زیر یک درخت تصمیم است:

رمزگشا

#زبان

به طور کلی ، هر سیستم ML که از یک بازنمایی فرآوری شده ، متراکم یا داخلی به نمایندگی خام تر ، پراکنده تر یا خارجی تبدیل می شود.

رمزگشایی ها اغلب جزء یک مدل بزرگتر هستند ، جایی که اغلب با یک رمزگذار جفت می شوند.

در کارهای توالی به توالی ، یک رمزگشایی با وضعیت داخلی تولید شده توسط رمزگذار شروع می شود تا دنباله بعدی را پیش بینی کند.

برای تعریف رمزگشایی در معماری ترانسفورماتور به ترانسفورماتور مراجعه کنید.

مدل عمیق

#فونداستال ها

یک شبکه عصبی حاوی بیش از یک لایه پنهان .

یک مدل عمیق نیز یک شبکه عصبی عمیق نامیده می شود.

تضاد با مدل گسترده .

شبکه عصبی عمیق

مترادف برای مدل عمیق .

شبکه Q Deep (DQN)

#RL

در یادگیری Q ، یک شبکه عصبی عمیق که عملکرد Q را پیش بینی می کند.

منتقد مترادف برای شبکه Q Deep است.

برابری جمعیتی

#فریبی

متریک انصاف که اگر نتایج طبقه بندی یک مدل به یک ویژگی حساس خاص وابسته نباشد ، راضی است.

به عنوان مثال ، اگر هر دو لیلیپوتیایی و Brobdingnagians در دانشگاه Glubbdubdrib اعمال شوند ، در صورتی که درصد لیلیپوتیان پذیرفته شده همانند درصد BrobdingNagians اعتراف کند ، بدون توجه به اینکه یک گروه به طور متوسط ​​نسبت به گروه دیگر واجد شرایط تر هستند ، برابری جمعیتی حاصل می شود.

تضاد با شانس مساوی و برابری فرصت ، که باعث می شود طبقه بندی منجر به وابستگی به ویژگی های حساس شود ، اما نتایج طبقه بندی را برای برخی از برچسب های مشخص شده با حقیقت زمین به ویژگی های حساس اجازه نمی دهد. برای تجسم در مورد تجارت در هنگام بهینه سازی برای برابری جمعیتی ، به "حمله به تبعیض با یادگیری ماشین هوشمند" مراجعه کنید.

حذف نویز

#زبان

یک رویکرد مشترک برای یادگیری خودبوشی که در آن:

1. نویز به صورت مصنوعی به مجموعه داده اضافه می شود.
2. مدل سعی می کند نویز را از بین ببرد.

Denoising یادگیری از نمونه های بدون برچسب را امکان پذیر می کند. مجموعه داده اصلی به عنوان هدف یا برچسب و داده های پر سر و صدا به عنوان ورودی عمل می کند.

برخی از مدل های زبانی نقاب دار از Denoising به شرح زیر استفاده می کنند:

1. سر و صدا با ماسک کردن برخی از نشانه ها به یک جمله بدون برچسب اضافه می شود.
2. این مدل سعی دارد نشانه های اصلی را پیش بینی کند.

ویژگی متراکم

#فونداستال ها

ویژگی ای که در آن بیشتر یا تمام مقادیر nonzero هستند ، به طور معمول یک تانسور از مقادیر نقطه شناور. به عنوان مثال ، تانسور 10 عنصر زیر متراکم است زیرا 9 از مقادیر آن Nonzero هستند:

8

3

7

5

2

4

0

4

9

6

تضاد با ویژگی پراکنده .

لایه متراکم

مترادف برای لایه کاملاً متصل .

عمق

#فونداستال ها

مبلغ موارد زیر در یک شبکه عصبی :

• تعداد لایه های پنهان
• تعداد لایه های خروجی ، که به طور معمول 1 است
• تعداد هر لایه تعبیه شده

به عنوان مثال ، یک شبکه عصبی با پنج لایه پنهان و یک لایه خروجی عمق 6 دارد.

توجه کنید که لایه ورودی بر عمق تأثیر نمی گذارد.

شبکه عصبی Convolutional Convolutional Depthwise (SEPCNN)

#تصویر

یک معماری شبکه عصبی حلقوی مبتنی بر شروع ، اما در جایی که ماژول های آغازین با پیچیدگی های جدا از عمق جایگزین می شوند. همچنین به عنوان Xception شناخته می شود.

یک همبستگی جدا از عمق (که به عنوان همبستگی قابل جدا شدن نیز به طور خلاصه) یک همبستگی 3 بعدی استاندارد را به دو عملیات جداگانه جداگانه که از نظر محاسباتی کارآمدتر هستند ، عوامل دارد: ابتدا ، یک همبستگی عمق ، با عمق 1 (n ✕ n ✕ 1) ، و سپس دوم ، یک حلقوی نقطه ای ، با طول و عرض 1 (1 ✕ 1 ✕ N).

برای کسب اطلاعات بیشتر ، به Xception: Learning Deep با پیچیدگی های جداگانه قابل جدا شدن مراجعه کنید.

برچسب مشتق شده

مترادف برچسب پروکسی .

دستگاه

#tensorflow

#GOOGLECLOUD

یک اصطلاح بیش از حد با دو تعریف ممکن زیر:

1. دسته سخت افزاری که می تواند یک جلسه Tensorflow از جمله CPU ، GPU و TPU را اجرا کند.
2. هنگام آموزش یک مدل ML در تراشه های شتاب دهنده (GPU یا TPU) ، بخشی از سیستم که در واقع تانسور و تعبیه را دستکاری می کند. دستگاه روی تراشه های شتاب دهنده اجرا می شود. در مقابل ، میزبان به طور معمول روی یک پردازنده اجرا می شود.

حریم خصوصی

یک رویکرد ناشناس سازی به حریم خصوصی که از اطلاعات شخصی یک فرد محافظت می کند که ممکن است در مجموعه آموزش یک مدل گنجانده شود. این رویکرد تضمین می کند که مدل در مورد یک فرد خاص استنباط نمی کند. حریم خصوصی دیفرانسیل در حین آموزش سر و صدا تزریق می کند تا نقاط داده های فردی را مبهم کند.

حریم خصوصی دیفرانسیل نیز در خارج از یادگیری ماشین استفاده می شود. به عنوان مثال ، دانشمندان داده ها گاهی اوقات از حریم خصوصی دیفرانسیل برای محافظت از حریم شخصی فردی هنگام محاسبه آمار استفاده از محصول برای جمعیتی مختلف استفاده می کنند.

کاهش بعد

Decreasing the number of dimensions used to represent a particular feature in a feature vector, typically by converting to an embedding vector .

ابعاد

Overloaded term having any of the following definitions:

• The number of levels of coordinates in a Tensor . مثلا:

  • A scalar has zero dimensions; for example, ["Hello"] .
  • A vector has one dimension; for example, [3, 5, 7, 11] .
  • A matrix has two dimensions; for example, [[2, 4, 18], [5, 7, 14]] .

  You can uniquely specify a particular cell in a one-dimensional vector with one coordinate; you need two coordinates to uniquely specify a particular cell in a two-dimensional matrix.

• The number of entries in a feature vector .

• The number of elements in an embedding layer .

direct prompting

#زبان

#generativeAI

Synonym for zero-shot prompting .

discrete feature

#fundamentals

A feature with a finite set of possible values. For example, a feature whose values may only be animal , vegetable , or mineral is a discrete (or categorical) feature.

Contrast with continuous feature .

discriminative model

A model that predicts labels from a set of one or more features . More formally, discriminative models define the conditional probability of an output given the features and weights ; به این معنا که:

p(output | features, weights)

For example, a model that predicts whether an email is spam from features and weights is a discriminative model.

The vast majority of supervised learning models, including classification and regression models, are discriminative models.

Contrast with generative model .

تبعیض کننده

A system that determines whether examples are real or fake.

Alternatively, the subsystem within a generative adversarial network that determines whether the examples created by the generator are real or fake.

تاثیر متفاوت

#fairness

Making decisions about people that impact different population subgroups disproportionately. This usually refers to situations where an algorithmic decision-making process harms or benefits some subgroups more than others.

For example, suppose an algorithm that determines a Lilliputian's eligibility for a miniature-home loan is more likely to classify them as “ineligible” if their mailing address contains a certain postal code. If Big-Endian Lilliputians are more likely to have mailing addresses with this postal code than Little-Endian Lilliputians, then this algorithm may result in disparate impact.

Contrast with disparate treatment , which focuses on disparities that result when subgroup characteristics are explicit inputs to an algorithmic decision-making process.

disparate treatment

#fairness

Factoring subjects' sensitive attributes into an algorithmic decision-making process such that different subgroups of people are treated differently.

For example, consider an algorithm that determines Lilliputians' eligibility for a miniature-home loan based on the data they provide in their loan application. If the algorithm uses a Lilliputian's affiliation as Big-Endian or Little-Endian as an input, it is enacting disparate treatment along that dimension.

Contrast with disparate impact , which focuses on disparities in the societal impacts of algorithmic decisions on subgroups, irrespective of whether those subgroups are inputs to the model.

Warning: Because sensitive attributes are almost always correlated with other features the data may have, explicitly removing sensitive attribute information doesn't guarantee that subgroups will be treated equally. For example, removing sensitive demographic attributes from a training data set that still includes postal code as a feature may address disparate treatment of subgroups, but there still might be disparate impact upon these groups because postal code might serve as a proxy for other demographic information.

تقطیر

#generativeAI

The process of reducing the size of one model (known as the teacher ) into a smaller model (known as the student ) that emulates the original model's predictions as faithfully as possible. Distillation is useful because the smaller model has two key benefits over the larger model (the teacher):

• Faster inference time
• Reduced memory and energy usage

However, the student's predictions are typically not as good as the teacher's predictions.

Distillation trains the student model to minimize a loss function based on the difference between the outputs of the predictions of the student and teacher models.

Compare and contrast distillation with the following terms:

• تنظیم دقیق
• prompt-based learning

divisive clustering

#clustering

See hierarchical clustering .

downsampling

#image

Overloaded term that can mean either of the following:

• Reducing the amount of information in a feature in order to train a model more efficiently. For example, before training an image recognition model, downsampling high-resolution images to a lower-resolution format.
• Training on a disproportionately low percentage of over-represented class examples in order to improve model training on under-represented classes. For example, in a class-imbalanced dataset , models tend to learn a lot about the majority class and not enough about the minority class . Downsampling helps balance the amount of training on the majority and minority classes.

DQN

#rl

Abbreviation for Deep Q-Network .

dropout regularization

A form of regularization useful in training neural networks . Dropout regularization removes a random selection of a fixed number of the units in a network layer for a single gradient step. The more units dropped out, the stronger the regularization. This is analogous to training the network to emulate an exponentially large ensemble of smaller networks. For full details, see Dropout: A Simple Way to Prevent Neural Networks from Overfitting .

پویا

#fundamentals

Something done frequently or continuously. The terms dynamic and online are synonyms in machine learning. The following are common uses of dynamic and online in machine learning:

• A dynamic model (or online model ) is a model that is retrained frequently or continuously.
• Dynamic training (or online training ) is the process of training frequently or continuously.
• Dynamic inference (or online inference ) is the process of generating predictions on demand.

dynamic model

#fundamentals

A model that is frequently (maybe even continuously) retrained. A dynamic model is a "lifelong learner" that constantly adapts to evolving data. A dynamic model is also known as an online model .

Contrast with static model .

E

eager execution

#TensorFlow

A TensorFlow programming environment in which operations run immediately. In contrast, operations called in graph execution don't run until they are explicitly evaluated. Eager execution is an imperative interface , much like the code in most programming languages. Eager execution programs are generally far easier to debug than graph execution programs.

توقف زودهنگام

#fundamentals

A method for regularization that involves ending training before training loss finishes decreasing. In early stopping, you intentionally stop training the model when the loss on a validation dataset starts to increase; that is, when generalization performance worsens.

Click the icon for additional notes.

Early stopping may seem counterintuitive. After all, telling a model to halt training while the loss is still decreasing may seem like telling a chef to stop cooking before the dessert has fully baked. However, training a model for too long can lead to overfitting . That is, if you train a model too long, the model may fit the training data so closely that the model doesn't make good predictions on new examples.

---

earth mover's distance (EMD)

A measure of the relative similarity between two documents. The lower the earth mover's distance, the more similar the documents.

edit distance

#زبان

A measurement of how similar two text strings are to each other. In machine learning, edit distance is useful because it is simple and easy to compute, and an effective way to compare two strings that are known to be similar or to find strings that are similar to a given string.

There are several definitions of edit distance, each using different string operations. For example, the Levenshtein distance considers the fewest delete, insert, and substitute operations.

For example, the Levenshtein distance between the words "heart" and "darts" is 3 because the following 3 edits are the fewest changes to turn one word into the other:

1. heart → deart (substitute “h” with "d")
2. deart → dart (delete "e")
3. dart → darts (insert "s")

Einsum notation

An efficient notation for describing how two tensors are to be combined. The tensors are combined by multiplying the elements of one tensor by the elements of the other tensor and then summing the products. Einsum notation uses symbols to identify the axes of each tensor, and those same symbols are rearranged to specify the shape of the new resulting tensor.

NumPy provides a common Einsum implementation.

embedding layer

#زبان

#fundamentals

A special hidden layer that trains on a high-dimensional categorical feature to gradually learn a lower dimension embedding vector. An embedding layer enables a neural network to train far more efficiently than training just on the high-dimensional categorical feature.

For example, Earth currently supports about 73,000 tree species. Suppose tree species is a feature in your model, so your model's input layer includes a one-hot vector 73,000 elements long. For example, perhaps baobab would be represented something like this:

A 73,000-element array is very long. If you don't add an embedding layer to the model, training is going to be very time consuming due to multiplying 72,999 zeros. Perhaps you pick the embedding layer to consist of 12 dimensions. Consequently, the embedding layer will gradually learn a new embedding vector for each tree species.

In certain situations, hashing is a reasonable alternative to an embedding layer.

embedding space

#زبان

The d-dimensional vector space that features from a higher-dimensional vector space are mapped to. Ideally, the embedding space contains a structure that yields meaningful mathematical results; for example, in an ideal embedding space, addition and subtraction of embeddings can solve word analogy tasks.

The dot product of two embeddings is a measure of their similarity.

embedding vector

#زبان

Broadly speaking, an array of floating-point numbers taken from any hidden layer that describe the inputs to that hidden layer. Often, an embedding vector is the array of floating-point numbers trained in an embedding layer. For example, suppose an embedding layer must learn an embedding vector for each of the 73,000 tree species on Earth. Perhaps the following array is the embedding vector for a baobab tree:

An embedding vector is not a bunch of random numbers. An embedding layer determines these values through training, similar to the way a neural network learns other weights during training. Each element of the array is a rating along some characteristic of a tree species. Which element represents which tree species' characteristic? That's very hard for humans to determine.

The mathematically remarkable part of an embedding vector is that similar items have similar sets of floating-point numbers. For example, similar tree species have a more similar set of floating-point numbers than dissimilar tree species. Redwoods and sequoias are related tree species, so they'll have a more similar set of floating-pointing numbers than redwoods and coconut palms. The numbers in the embedding vector will change each time you retrain the model, even if you retrain the model with identical input.

empirical risk minimization (ERM)

Choosing the function that minimizes loss on the training set. Contrast with structural risk minimization .

رمزگذار

#زبان

In general, any ML system that converts from a raw, sparse, or external representation into a more processed, denser, or more internal representation.

Encoders are often a component of a larger model, where they are frequently paired with a decoder . Some Transformers pair encoders with decoders, though other Transformers use only the encoder or only the decoder.

Some systems use the encoder's output as the input to a classification or regression network.

In sequence-to-sequence tasks , an encoder takes an input sequence and returns an internal state (a vector). Then, the decoder uses that internal state to predict the next sequence.

Refer to Transformer for the definition of an encoder in the Transformer architecture.

گروه

A collection of models trained independently whose predictions are averaged or aggregated. In many cases, an ensemble produces better predictions than a single model. For example, a random forest is an ensemble built from multiple decision trees . Note that not all decision forests are ensembles.

آنتروپی

#df

In information theory , a description of how unpredictable a probability distribution is. Alternatively, entropy is also defined as how much information each example contains. A distribution has the highest possible entropy when all values of a random variable are equally likely.

The entropy of a set with two possible values "0" and "1" (for example, the labels in a binary classification problem) has the following formula:

H = -p log p - q log q = -p log p - (1-p) * log (1-p)

جایی که:

• H is the entropy.
• p is the fraction of "1" examples.
• q is the fraction of "0" examples. Note that q = (1 - p)
• log is generally log ₂ . In this case, the entropy unit is a bit.

برای مثال موارد زیر را فرض کنید:

• 100 examples contain the value "1"
• 300 examples contain the value "0"

Therefore, the entropy value is:

• p = 0.25
• q = 0.75
• H = (-0.25)log ₂ (0.25) - (0.75)log ₂ (0.75) = 0.81
bits per example

A set that is perfectly balanced (for example, 200 "0"s and 200 "1"s) would have an entropy of 1.0 bit per example. As a set becomes more imbalanced , its entropy moves towards 0.0.

In decision trees , entropy helps formulate information gain to help the splitter select the conditions during the growth of a classification decision tree.

Compare entropy with:

• ناخالصی جینی
• cross-entropy loss function

Entropy is often called Shannon's entropy.

محیط

#rl

In reinforcement learning, the world that contains the agent and allows the agent to observe that world's state . For example, the represented world can be a game like chess, or a physical world like a maze. When the agent applies an action to the environment, then the environment transitions between states.

قسمت

#rl

In reinforcement learning, each of the repeated attempts by the agent to learn an environment .

دوران

#fundamentals

A full training pass over the entire training set such that each example has been processed once.

An epoch represents N / batch size training iterations , where N is the total number of examples.

For instance, suppose the following:

• The dataset consists of 1,000 examples.
• The batch size is 50 examples.

Therefore, a single epoch requires 20 iterations:

1 epoch = (N/batch size) = (1,000 / 50) = 20 iterations

epsilon greedy policy

#rl

In reinforcement learning, a policy that either follows a random policy with epsilon probability or a greedy policy otherwise. For example, if epsilon is 0.9, then the policy follows a random policy 90% of the time and a greedy policy 10% of the time.

Over successive episodes, the algorithm reduces epsilon's value in order to shift from following a random policy to following a greedy policy. By shifting the policy, the agent first randomly explores the environment and then greedily exploits the results of random exploration.

برابری فرصت ها

#fairness

A fairness metric that checks whether, for a preferred label (one that confers an advantage or benefit to a person) and a given attribute , a classifier predicts that preferred label equally well for all values of that attribute. In other words, equality of opportunity measures whether the people who should qualify for an opportunity are equally likely to do so regardless of their group membership.

For example, suppose Glubbdubdrib University admits both Lilliputians and Brobdingnagians to a rigorous mathematics program. Lilliputians' secondary schools offer a robust curriculum of math classes, and the vast majority of students are qualified for the university program. Brobdingnagians' secondary schools don't offer math classes at all, and as a result, far fewer of their students are qualified. Equality of opportunity is satisfied for the preferred label of "admitted" with respect to nationality (Lilliputian or Brobdingnagian) if qualified students are equally likely to be admitted irrespective of whether they're a Lilliputian or a Brobdingnagian.

For example, let's say 100 Lilliputians and 100 Brobdingnagians apply to Glubbdubdrib University, and admissions decisions are made as follows:

Table 1. Lilliputian applicants (90% are qualified)

	واجد شرایط	Unqualified
پذیرفته	45	3
رد شد	45	7
جمع	90	10
Percentage of qualified students admitted: 45/90 = 50% Percentage of unqualified students rejected: 7/10 = 70% Total percentage of Lilliputian students admitted: (45+3)/100 = 48%

Table 2. Brobdingnagian applicants (10% are qualified):

	واجد شرایط	Unqualified
پذیرفته	5	9
رد شد	5	81
جمع	10	90
Percentage of qualified students admitted: 5/10 = 50% Percentage of unqualified students rejected: 81/90 = 90% Total percentage of Brobdingnagian students admitted: (5+9)/100 = 14%

The preceding examples satisfy equality of opportunity for acceptance of qualified students because qualified Lilliputians and Brobdingnagians both have a 50% chance of being admitted.

Note: While equality of opportunity is satisfied, the following two fairness metrics are not satisfied:

• demographic parity : Lilliputians and Brobdingnagians are admitted to the university at different rates; 48% of Lilliputians students are admitted, but only 14% of Brobdingnagian students are admitted.
• equalized odds : While qualified Lilliputian and Brobdingnagian students both have the same chance of being admitted, the additional constraint that unqualified Lilliputians and Brobdingnagians both have the same chance of being rejected is not satisfied. Unqualified Lilliputians have a 70% rejection rate, whereas unqualified Brobdingnagians have a 90% rejection rate.

See "Equality of Opportunity in Supervised Learning" for a more detailed discussion of equality of opportunity. Also see "Attacking discrimination with smarter machine learning" for a visualization exploring the tradeoffs when optimizing for equality of opportunity.

equalized odds

#fairness

A fairness metric that checks if, for any particular label and attribute, a classifier predicts that label equally well for all values of that attribute.

For example, suppose Glubbdubdrib University admits both Lilliputians and Brobdingnagians to a rigorous mathematics program. Lilliputians' secondary schools offer a robust curriculum of math classes, and the vast majority of students are qualified for the university program. Brobdingnagians' secondary schools don't offer math classes at all, and as a result, far fewer of their students are qualified. Equalized odds is satisfied provided that no matter whether an applicant is a Lilliputian or a Brobdingnagian, if they are qualified, they are equally as likely to get admitted to the program, and if they are not qualified, they are equally as likely to get rejected .

Let's say 100 Lilliputians and 100 Brobdingnagians apply to Glubbdubdrib University, and admissions decisions are made as follows:

Table 3. Lilliputian applicants (90% are qualified)

	واجد شرایط	Unqualified
پذیرفته	45	2
رد شد	45	8
جمع	90	10
Percentage of qualified students admitted: 45/90 = 50% Percentage of unqualified students rejected: 8/10 = 80% Total percentage of Lilliputian students admitted: (45+2)/100 = 47%

Table 4. Brobdingnagian applicants (10% are qualified):

	واجد شرایط	Unqualified
پذیرفته	5	18
رد شد	5	72
جمع	10	90
Percentage of qualified students admitted: 5/10 = 50% Percentage of unqualified students rejected: 72/90 = 80% Total percentage of Brobdingnagian students admitted: (5+18)/100 = 23%

Equalized odds is satisfied because qualified Lilliputian and Brobdingnagian students both have a 50% chance of being admitted, and unqualified Lilliputian and Brobdingnagian have an 80% chance of being rejected.

Note: While equalized odds is satisfied here, demographic parity is not satisfied . Lilliputian and Brobdingnagian students are admitted to Glubbdubdrib University at different rates; 47% of Lilliputian students are admitted, and 23% of Brobdingnagian students are admitted.

Equalized odds is formally defined in "Equality of Opportunity in Supervised Learning" as follows: "predictor Ŷ satisfies equalized odds with respect to protected attribute A and outcome Y if Ŷ and A are independent, conditional on Y."

Note: Contrast equalized odds with the more relaxed equality of opportunity metric.

برآوردگر

#TensorFlow

A deprecated TensorFlow API. Use tf.keras instead of Estimators.

مثال

#fundamentals

The values of one row of features and possibly a label . Examples in supervised learning fall into two general categories:

• A labeled example consists of one or more features and a label. Labeled examples are used during training.
• An unlabeled example consists of one or more features but no label. Unlabeled examples are used during inference.

For instance, suppose you are training a model to determine the influence of weather conditions on student test scores. Here are three labeled examples:

امکانات			برچسب
درجه حرارت	رطوبت	فشار	نمره آزمون
15	47	998	خوب
19	34	1020	عالی
18	92	1012	فقیر

Here are three unlabeled examples:

درجه حرارت	رطوبت	فشار
12	62	1014
21	47	1017
19	41	1021

The row of a dataset is typically the raw source for an example. That is, an example typically consists of a subset of the columns in the dataset. Furthermore, the features in an example can also include synthetic features , such as feature crosses .

experience replay

#rl

In reinforcement learning, a DQN technique used to reduce temporal correlations in training data. The agent stores state transitions in a replay buffer , and then samples transitions from the replay buffer to create training data.

experimenter's bias

#fairness

See confirmation bias .

exploding gradient problem

#seq

The tendency for gradients in deep neural networks (especially recurrent neural networks ) to become surprisingly steep (high). Steep gradients often cause very large updates to the weights of each node in a deep neural network.

Models suffering from the exploding gradient problem become difficult or impossible to train. Gradient clipping can mitigate this problem.

Compare to vanishing gradient problem .

اف

F ₁

A "roll-up" binary classification metric that relies on both precision and recall . این فرمول است:

$$F{_1} = \frac{\text{2 * precision * recall}} {\text{precision + recall}}$$

For example, given the following:

• precision = 0.6
• recall = 0.4

$$F{_1} = \frac{\text{2 * 0.6 * 0.4}} {\text{0.6 + 0.4}} = 0.48$$

When precision and recall are fairly similar (as in the preceding example), F ₁ is close to their mean. When precision and recall differ significantly, F ₁ is closer to the lower value. مثلا:

• precision = 0.9
• recall = 0.1

$$F{_1} = \frac{\text{2 * 0.9 * 0.1}} {\text{0.9 + 0.1}} = 0.18$$

fairness constraint

#fairness

Applying a constraint to an algorithm to ensure one or more definitions of fairness are satisfied. Examples of fairness constraints include:

• Post-processing your model's output.
• Altering the loss function to incorporate a penalty for violating a fairness metric .
• Directly adding a mathematical constraint to an optimization problem.

fairness metric

#fairness

A mathematical definition of “fairness” that is measurable. Some commonly used fairness metrics include:

• equalized odds
• predictive parity
• counterfactual fairness
• demographic parity

Many fairness metrics are mutually exclusive; see incompatibility of fairness metrics .

false negative (FN)

#fundamentals

An example in which the model mistakenly predicts the negative class . For example, the model predicts that a particular email message is not spam (the negative class), but that email message actually is spam .

false negative rate

The proportion of actual positive examples for which the model mistakenly predicted the negative class. The following formula calculates the false negative rate:

$$\text{false negative rate} = \frac{\text{false negatives}}{\text{false negatives} + \text{true positives}}$$

false positive (FP)

#fundamentals

An example in which the model mistakenly predicts the positive class . For example, the model predicts that a particular email message is spam (the positive class), but that email message is actually not spam .

false positive rate (FPR)

#fundamentals

The proportion of actual negative examples for which the model mistakenly predicted the positive class. The following formula calculates the false positive rate:

$$\text{false positive rate} = \frac{\text{false positives}}{\text{false positives} + \text{true negatives}}$$

The false positive rate is the x-axis in an ROC curve .

ویژگی

#fundamentals

An input variable to a machine learning model. An example consists of one or more features. For instance, suppose you are training a model to determine the influence of weather conditions on student test scores. The following table shows three examples, each of which contains three features and one label:

امکانات			برچسب
درجه حرارت	رطوبت	فشار	نمره آزمون
15	47	998	92
19	34	1020	84
18	92	1012	87

Contrast with label .

feature cross

#fundamentals

A synthetic feature formed by "crossing" categorical or bucketed features.

For example, consider a "mood forecasting" model that represents temperature in one of the following four buckets:

• freezing
• chilly
• temperate
• warm

And represents wind speed in one of the following three buckets:

• still
• light
• windy

Without feature crosses, the linear model trains independently on each of the preceding seven various buckets. So, the model trains on, for instance, freezing independently of the training on, for instance, windy .

Alternatively, you could create a feature cross of temperature and wind speed. This synthetic feature would have the following 12 possible values:

• freezing-still
• freezing-light
• freezing-windy
• chilly-still
• chilly-light
• chilly-windy
• temperate-still
• temperate-light
• temperate-windy
• warm-still
• warm-light
• warm-windy

Thanks to feature crosses, the model can learn mood differences between a freezing-windy day and a freezing-still day.

If you create a synthetic feature from two features that each have a lot of different buckets, the resulting feature cross will have a huge number of possible combinations. For example, if one feature has 1,000 buckets and the other feature has 2,000 buckets, the resulting feature cross has 2,000,000 buckets.

Formally, a cross is a Cartesian product .

Feature crosses are mostly used with linear models and are rarely used with neural networks.

feature engineering

#fundamentals

#TensorFlow

A process that involves the following steps:

1. Determining which features might be useful in training a model.
2. Converting raw data from the dataset into efficient versions of those features.

For example, you might determine that temperature might be a useful feature. Then, you might experiment with bucketing to optimize what the model can learn from different temperature ranges.

Feature engineering is sometimes called feature extraction .

Click the icon for additional notes about TensorFlow.

In TensorFlow, feature engineering often means converting raw log file entries to tf.Example protocol buffers. See also tf.Transform .

---

استخراج ویژگی

Overloaded term having either of the following definitions:

• Retrieving intermediate feature representations calculated by an unsupervised or pretrained model (for example, hidden layer values in a neural network ) for use in another model as input.
• Synonym for feature engineering .

feature importances

#df

Synonym for variable importances .

feature set

#fundamentals

The group of features your machine learning model trains on. For example, postal code, property size, and property condition might comprise a simple feature set for a model that predicts housing prices.

feature spec

#TensorFlow

Describes the information required to extract features data from the tf.Example protocol buffer. Because the tf.Example protocol buffer is just a container for data, you must specify the following:

• the data to extract (that is, the keys for the features)
• the data type (for example, float or int)
• The length (fixed or variable)

feature vector

#fundamentals

The array of feature values comprising an example . The feature vector is input during training and during inference . For example, the feature vector for a model with two discrete features might be:

[0.92, 0.56]

Each example supplies different values for the feature vector, so the feature vector for the next example could be something like:

[0.73, 0.49]

Feature engineering determines how to represent features in the feature vector. For example, a binary categorical feature with five possible values might be represented with one-hot encoding . In this case, the portion of the feature vector for a particular example would consist of four zeroes and a single 1.0 in the third position, as follows:

[0.0, 0.0, 1.0, 0.0, 0.0]

As another example, suppose your model consists of three features:

• a binary categorical feature with five possible values represented with one-hot encoding; for example: [0.0, 1.0, 0.0, 0.0, 0.0]
• another binary categorical feature with three possible values represented with one-hot encoding; for example: [0.0, 0.0, 1.0]
• a floating-point feature; for example: 8.3 .

In this case, the feature vector for each example would be represented by nine values. Given the example values in the preceding list, the feature vector would be:

0.0
1.0
0.0
0.0
0.0
0.0
0.0
1.0
8.3

یادگیری فدرال

A distributed machine learning approach that trains machine learning models using decentralized examples residing on devices such as smartphones. In federated learning, a subset of devices downloads the current model from a central coordinating server. The devices use the examples stored on the devices to make improvements to the model. The devices then upload the model improvements (but not the training examples) to the coordinating server, where they are aggregated with other updates to yield an improved global model. After the aggregation, the model updates computed by devices are no longer needed, and can be discarded.

Since the training examples are never uploaded, federated learning follows the privacy principles of focused data collection and data minimization.

For more information about federated learning, see this tutorial .

حلقه بازخورد

#fundamentals

In machine learning, a situation in which a model's predictions influence the training data for the same model or another model. For example, a model that recommends movies will influence the movies that people see, which will then influence subsequent movie recommendation models.

feedforward neural network (FFN)

A neural network without cyclic or recursive connections. For example, traditional deep neural networks are feedforward neural networks. Contrast with recurrent neural networks , which are cyclic.

few-shot learning

A machine learning approach, often used for object classification, designed to train effective classifiers from only a small number of training examples.

See also one-shot learning and zero-shot learning .

few-shot prompting

#زبان

#generativeAI

A prompt that contains more than one (a "few") example demonstrating how the large language model should respond. For example, the following lengthy prompt contains two examples showing a large language model how to answer a query.

Parts of one prompt	یادداشت
What is the official currency of the specified country?	The question you want the LLM to answer.
France: EUR	یک مثال.
United Kingdom: GBP	مثالی دیگر.
هند:	The actual query.

Few-shot prompting generally produces more desirable results than zero-shot prompting and one-shot prompting . However, few-shot prompting requires a lengthier prompt.

Few-shot prompting is a form of few-shot learning applied to prompt-based learning .

کمانچه

#زبان

A Python-first configuration library that sets the values of functions and classes without invasive code or infrastructure. In the case of Pax —and other ML codebases—these functions and classes represent models and training hyperparameters .

Fiddle assumes that machine learning codebases are typically divided into:

• Library code, which defines the layers and optimizers.
• Dataset "glue" code, which calls the libraries and wires everything together.

Fiddle captures the call structure of the glue code in an unevaluated and mutable form.

تنظیم دقیق

#زبان

#image

#generativeAI

A second, task-specific training pass performed on a pre-trained model to refine its parameters for a specific use case. For example, the full training sequence for some large language models is as follows:

1. Pre-training: Train a large language model on a vast general dataset, such as all the English language Wikipedia pages.
2. Fine-tuning: Train the pre-trained model to perform a specific task, such as responding to medical queries. Fine-tuning typically involves hundreds or thousands of examples focused on the specific task.

As another example, the full training sequence for a large image model is as follows:

1. Pre-training: Train a large image model on a vast general image dataset, such as all the images in Wikimedia commons.
2. Fine-tuning: Train the pre-trained model to perform a specific task, such as generating images of orcas.

Fine-tuning can entail any combination of the following strategies:

• Modifying all of the pre-trained model's existing parameters . This is sometimes called full fine-tuning .
• Modifying only some of the pre-trained model's existing parameters (typically, the layers closest to the output layer ), while keeping other existing parameters unchanged (typically, the layers closest to the input layer ). See parameter-efficient tuning .
• Adding more layers, typically on top of the existing layers closest to the output layer.

Fine-tuning is a form of transfer learning . As such, fine-tuning might use a different loss function or a different model type than those used to train the pre-trained model. For example, you could fine-tune a pre-trained large image model to produce a regression model that returns the number of birds in an input image.

Compare and contrast fine-tuning with the following terms:

• تقطیر
• prompt-based learning

کتان

#زبان

A high-performance open-source library for deep learning built on top of JAX . Flax provides functions for training neural networks , as well as methods for evaluating their performance.

Flaxformer

#زبان

An open-source Transformer library , built on Flax , designed primarily for natural language processing and multimodal research.

forget gate

#seq

The portion of a Long Short-Term Memory cell that regulates the flow of information through the cell. Forget gates maintain context by deciding which information to discard from the cell state.

full softmax

Synonym for softmax .

Contrast with candidate sampling .

لایه کاملا متصل

A hidden layer in which each node is connected to every node in the subsequent hidden layer.

A fully connected layer is also known as a dense layer .

جی

GAN

Abbreviation for generative adversarial network .

تعمیم

#fundamentals

A model's ability to make correct predictions on new, previously unseen data. A model that can generalize is the opposite of a model that is overfitting .

Click the icon for additional notes.

You train a model on the examples in the training set. Consequently, the model learns the peculiarities of the data in the training set. Generalization essentially asks whether your model can make good predictions on examples that are not in the training set.

To encourage generalization, regularization helps a model train less exactly to the peculiarities of the data in the training set.

---

generalization curve

#fundamentals

A plot of both training loss and validation loss as a function of the number of iterations .

A generalization curve can help you detect possible overfitting . For example, the following generalization curve suggests overfitting because validation loss ultimately becomes significantly higher than training loss.

generalized linear model

A generalization of least squares regression models, which are based on Gaussian noise , to other types of models based on other types of noise, such as Poisson noise or categorical noise. Examples of generalized linear models include:

• رگرسیون لجستیک
• multi-class regression
• رگرسیون حداقل مربعات

The parameters of a generalized linear model can be found through convex optimization .

Generalized linear models exhibit the following properties:

• The average prediction of the optimal least squares regression model is equal to the average label on the training data.
• The average probability predicted by the optimal logistic regression model is equal to the average label on the training data.

The power of a generalized linear model is limited by its features. Unlike a deep model, a generalized linear model cannot "learn new features."

شبکه متخاصم مولد (GAN)

A system to create new data in which a generator creates data and a discriminator determines whether that created data is valid or invalid.

هوش مصنوعی مولد

#زبان

#image

#generativeAI

An emerging transformative field with no formal definition. That said, most experts agree that generative AI models can create ("generate") content that is all of the following:

• مجتمع
• منسجم
• اصلی

For example, a generative AI model can create sophisticated essays or images.

Some earlier technologies, including LSTMs and RNNs , can also generate original and coherent content. Some experts view these earlier technologies as generative AI, while others feel that true generative AI requires more complex output than those earlier technologies can produce.

Contrast with predictive ML .

generative model

Practically speaking, a model that does either of the following:

• Creates (generates) new examples from the training dataset. For example, a generative model could create poetry after training on a dataset of poems. The generator part of a generative adversarial network falls into this category.
• Determines the probability that a new example comes from the training set, or was created from the same mechanism that created the training set. For example, after training on a dataset consisting of English sentences, a generative model could determine the probability that new input is a valid English sentence.

A generative model can theoretically discern the distribution of examples or particular features in a dataset. به این معنا که:

p(examples)

Unsupervised learning models are generative.

Contrast with discriminative models .

ژنراتور

The subsystem within a generative adversarial network that creates new examples .

Contrast with discriminative model .

ناخالصی جینی

#df

A metric similar to entropy . Splitters use values derived from either gini impurity or entropy to compose conditions for classification decision trees . Information gain is derived from entropy. There is no universally accepted equivalent term for the metric derived from gini impurity; however, this unnamed metric is just as important as information gain.

Gini impurity is also called gini index , or simply gini .

Click the icon for mathematical details about gini impurity.

Gini impurity is the probability of misclassifying a new piece of data taken from the same distribution. The gini impurity of a set with two possible values "0" and "1" (for example, the labels in a binary classification problem) is calculated from the following formula:

I = 1 - (p ² + q ² ) = 1 - (p ² + (1-p) ² )

جایی که:

• I is the gini impurity.
• p is the fraction of "1" examples.
• q is the fraction of "0" examples. Note that q = 1-p

For example, consider the following dataset:

• 100 labels (0.25 of the dataset) contain the value "1"
• 300 labels (0.75 of the dataset) contain the value "0"

Therefore, the gini impurity is:

p = 0.25

q = 0.75

I = 1 - (0.25 ² + 0.75 ² ) = 0.375

Consequently, a random label from the same dataset would have a 37.5% chance of being misclassified, and a 62.5% chance of being properly classified.

A perfectly balanced label (for example, 200 "0"s and 200 "1"s) would have a gini impurity of 0.5. A highly imbalanced label would have a gini impurity close to 0.0.

---

GPT (Generative Pre-trained Transformer)

#زبان

A family of Transformer -based large language models developed by OpenAI .

GPT variants can apply to multiple modalities , including:

• image generation (for example, ImageGPT)
• text-to-image generation (for example, DALL-E ).

شیب

The vector of partial derivatives with respect to all of the independent variables. In machine learning, the gradient is the vector of partial derivatives of the model function. The gradient points in the direction of steepest ascent.

gradient boosted (decision) trees (GBT)

#df

A type of decision forest in which:

• Training relies on gradient boosting .
• The weak model is a decision tree .

gradient boosting

#df

A training algorithm where weak models are trained to iteratively improve the quality (reduce the loss) of a strong model. For example, a weak model could be a linear or small decision tree model. The strong model becomes the sum of all the previously trained weak models.

In the simplest form of gradient boosting, at each iteration, a weak model is trained to predict the loss gradient of the strong model. Then, the strong model's output is updated by subtracting the predicted gradient, similar to gradient descent .

$$F_{0} = 0$$$$F_{i+1} = F_i - \xi f_i $$

جایی که:

• $F_{0}$ is the starting strong model.
• $F_{i+1}$ is the next strong model.
• $F_{i}$ is the current strong model.
• $\xi$ is a value between 0.0 and 1.0 called shrinkage , which is analogous to the learning rate in gradient descent.
• $f_{i}$ is the weak model trained to predict the loss gradient of $F_{i}$.

Modern variations of gradient boosting also include the second derivative (Hessian) of the loss in their computation.

Decision trees are commonly used as weak models in gradient boosting. See gradient boosted (decision) trees .

gradient clipping

#seq

A commonly used mechanism to mitigate the exploding gradient problem by artificially limiting (clipping) the maximum value of gradients when using gradient descent to train a model.

gradient descent

#fundamentals

A mathematical technique to minimize loss . Gradient descent iteratively adjusts weights and biases , gradually finding the best combination to minimize loss.

Gradient descent is older—much, much older—than machine learning.

نمودار

#TensorFlow

In TensorFlow, a computation specification. Nodes in the graph represent operations. Edges are directed and represent passing the result of an operation (a Tensor ) as an operand to another operation. Use TensorBoard to visualize a graph.

graph execution

#TensorFlow

A TensorFlow programming environment in which the program first constructs a graph and then executes all or part of that graph. Graph execution is the default execution mode in TensorFlow 1.x.

Contrast with eager execution .

greedy policy

#rl

In reinforcement learning, a policy that always chooses the action with the highest expected return .

ground truth

#fundamentals

واقعیت.

The thing that actually happened.

For example, consider a binary classification model that predicts whether a student in their first year of university will graduate within six years. Ground truth for this model is whether or not that student actually graduated within six years.

Click the icon for additional notes.

We assess model quality against ground truth. However, ground truth is not always completely, well, truthful. For example, consider the following examples of potential imperfections in ground truth:

• In the graduation example, are we certain that the graduation records for each student are always correct? Is the university's record-keeping flawless?
• Suppose the label is a floating-point value measured by instruments (for instance, barometers). How can we be sure that each instrument is calibrated identically or that each reading was taken under the same circumstances?
• If the label is a matter of human opinion, how can we be sure that each human rater is evaluating events in the same way? To improve consistency, expert human raters sometimes intervene.

---

group attribution bias

#fairness

Assuming that what is true for an individual is also true for everyone in that group. The effects of group attribution bias can be exacerbated if a convenience sampling is used for data collection. In a non-representative sample, attributions may be made that do not reflect reality.

See also out-group homogeneity bias and in-group bias .

اچ

توهم

#زبان

The production of plausible-seeming but factually incorrect output by a generative AI model that purports to be making an assertion about the real world. For example, a generative AI model that claims that Barack Obama died in 1865 is hallucinating .

هش کردن

In machine learning, a mechanism for bucketing categorical data , particularly when the number of categories is large, but the number of categories actually appearing in the dataset is comparatively small.

For example, Earth is home to about 73,000 tree species. You could represent each of the 73,000 tree species in 73,000 separate categorical buckets. Alternatively, if only 200 of those tree species actually appear in a dataset, you could use hashing to divide tree species into perhaps 500 buckets.

A single bucket could contain multiple tree species. For example, hashing could place baobab and red maple —two genetically dissimilar species—into the same bucket. Regardless, hashing is still a good way to map large categorical sets into the desired number of buckets. Hashing turns a categorical feature having a large number of possible values into a much smaller number of values by grouping values in a deterministic way.

ابتکاری

A simple and quickly implemented solution to a problem. For example, "With a heuristic, we achieved 86% accuracy. When we switched to a deep neural network, accuracy went up to 98%."

hidden layer

#fundamentals

A layer in a neural network between the input layer (the features) and the output layer (the prediction). Each hidden layer consists of one or more neurons . For example, the following neural network contains two hidden layers, the first with three neurons and the second with two neurons:

A deep neural network contains more than one hidden layer. For example, the preceding illustration is a deep neural network because the model contains two hidden layers.

خوشه بندی سلسله مراتبی

#clustering

A category of clustering algorithms that create a tree of clusters. Hierarchical clustering is well-suited to hierarchical data, such as botanical taxonomies. There are two types of hierarchical clustering algorithms:

• Agglomerative clustering first assigns every example to its own cluster, and iteratively merges the closest clusters to create a hierarchical tree.
• Divisive clustering first groups all examples into one cluster and then iteratively divides the cluster into a hierarchical tree.

Contrast with centroid-based clustering .

hinge loss

A family of loss functions for classification designed to find the decision boundary as distant as possible from each training example, thus maximizing the margin between examples and the boundary. KSVMs use hinge loss (or a related function, such as squared hinge loss). For binary classification, the hinge loss function is defined as follows:

$$\text{loss} = \text{max}(0, 1 - (y * y'))$$

where y is the true label, either -1 or +1, and y' is the raw output of the classifier model:

$$y' = b + w_1x_1 + w_2x_2 + … w_nx_n$$

Consequently, a plot of hinge loss vs. (y * y') looks as follows:

holdout data

Examples intentionally not used ("held out") during training. The validation dataset and test dataset are examples of holdout data. Holdout data helps evaluate your model's ability to generalize to data other than the data it was trained on. The loss on the holdout set provides a better estimate of the loss on an unseen dataset than does the loss on the training set.

میزبان

#TensorFlow

#GoogleCloud

When training an ML model on accelerator chips (GPUs or TPUs ), the part of the system that controls both of the following:

• The overall flow of the code.
• The extraction and transformation of the input pipeline.

The host typically runs on a CPU, not on an accelerator chip; the device manipulates tensors on the accelerator chips.

فرا پارامتر

#fundamentals

The variables that you or a hyperparameter tuning serviceadjust during successive runs of training a model. For example, learning rate is a hyperparameter. You could set the learning rate to 0.01 before one training session. If you determine that 0.01 is too high, you could perhaps set the learning rate to 0.003 for the next training session.

In contrast, parameters are the various weights and bias that the model learns during training.

ابر هواپیما

A boundary that separates a space into two subspaces. For example, a line is a hyperplane in two dimensions and a plane is a hyperplane in three dimensions. More typically in machine learning, a hyperplane is the boundary separating a high-dimensional space. Kernel Support Vector Machines use hyperplanes to separate positive classes from negative classes, often in a very high-dimensional space.

من

iid

Abbreviation for independently and identically distributed .

تشخیص تصویر

#image

A process that classifies object(s), pattern(s), or concept(s) in an image. Image recognition is also known as image classification .

For more information, see ML Practicum: Image Classification .

imbalanced dataset

Synonym for class-imbalanced dataset .

implicit bias

#fairness

Automatically making an association or assumption based on one's mental models and memories. Implicit bias can affect the following:

• How data is collected and classified.
• How machine learning systems are designed and developed.

For example, when building a classifier to identify wedding photos, an engineer may use the presence of a white dress in a photo as a feature. However, white dresses have been customary only during certain eras and in certain cultures.

See also confirmation bias .

انتساب

Short form of value imputation .

incompatibility of fairness metrics

#fairness

The idea that some notions of fairness are mutually incompatible and cannot be satisfied simultaneously. As a result, there is no single universal metric for quantifying fairness that can be applied to all ML problems.

While this may seem discouraging, incompatibility of fairness metrics doesn't imply that fairness efforts are fruitless. Instead, it suggests that fairness must be defined contextually for a given ML problem, with the goal of preventing harms specific to its use cases.

See "On the (im)possibility of fairness" for a more detailed discussion of this topic.

in-context learning

#زبان

#generativeAI

Synonym for few-shot prompting .

independently and identically distributed (iid)

#fundamentals

Data drawn from a distribution that doesn't change, and where each value drawn doesn't depend on values that have been drawn previously. An iid is the ideal gas of machine learning—a useful mathematical construct but almost never exactly found in the real world. For example, the distribution of visitors to a web page may be iid over a brief window of time; that is, the distribution doesn't change during that brief window and one person's visit is generally independent of another's visit. However, if you expand that window of time, seasonal differences in the web page's visitors may appear.

See also nonstationarity .

individual fairness

#fairness

A fairness metric that checks whether similar individuals are classified similarly. For example, Brobdingnagian Academy might want to satisfy individual fairness by ensuring that two students with identical grades and standardized test scores are equally likely to gain admission.

Note that individual fairness relies entirely on how you define "similarity" (in this case, grades and test scores), and you can run the risk of introducing new fairness problems if your similarity metric misses important information (such as the rigor of a student's برنامه تحصیلی).

See "Fairness Through Awareness" for a more detailed discussion of individual fairness.

استنتاج

#fundamentals

In machine learning, the process of making predictions by applying a trained model to unlabeled examples .

Inference has a somewhat different meaning in statistics. See the Wikipedia article on statistical inference for details.

inference path

#df

In a decision tree , during inference , the route a particular example takes from the root to other conditions , terminating with a leaf . For instance, in the following decision tree, the thicker arrows show the inference path for an example with the following feature values:

• x = 7
• y = 12
• z = -3

The inference path in the following illustration travels through three conditions before reaching the leaf ( Zeta ).

The three thick arrows show the inference path.

کسب اطلاعات

#df

In decision forests , the difference between a node's entropy and the weighted (by number of examples) sum of the entropy of its children nodes. A node's entropy is the entropy of the examples in that node.

For example, consider the following entropy values:

• entropy of parent node = 0.6
• entropy of one child node with 16 relevant examples = 0.2
• entropy of another child node with 24 relevant examples = 0.1

So 40% of the examples are in one child node and 60% are in the other child node. از این رو:

• weighted entropy sum of child nodes = (0.4 * 0.2) + (0.6 * 0.1) = 0.14

So, the information gain is:

• information gain = entropy of parent node - weighted entropy sum of child nodes
• information gain = 0.6 - 0.14 = 0.46

Most splitters seek to create conditions that maximize information gain.

in-group bias

#fairness

Showing partiality to one's own group or own characteristics. If testers or raters consist of the machine learning developer's friends, family, or colleagues, then in-group bias may invalidate product testing or the dataset.

In-group bias is a form of group attribution bias . See also out-group homogeneity bias .

input generator

A mechanism by which data is loaded into a neural network .

An input generator can be thought of as a component responsible for processing raw data into tensors which are iterated over to generate batches for training, evaluation, and inference.

input layer

#fundamentals

The layer of a neural network that holds the feature vector . That is, the input layer provides examples for training or inference . For example, the input layer in the following neural network consists of two features:

in-set condition

#df

In a decision tree , a condition that tests for the presence of one item in a set of items. For example, the following is an in-set condition:

  house-style in [tudor, colonial, cape]

During inference, if the value of the house-style feature is tudor or colonial or cape , then this condition evaluates to Yes. If the value of the house-style feature is something else (for example, ranch ), then this condition evaluates to No.

In-set conditions usually lead to more efficient decision trees than conditions that test one-hot encoded features.

نمونه، مثال

Synonym for example .

instruction tuning

#generativeAI

A form of fine-tuning that improves a generative AI model's ability to follow instructions. Instruction tuning involves training a model on a series of instruction prompts, typically covering a wide variety of tasks. The resulting instruction-tuned model then tends to generate useful responses to zero-shot prompts across a variety of tasks.

Compare and contrast with:

• parameter-efficient tuning
• prompt tuning

تفسیر پذیری

#fundamentals

The ability to explain or to present an ML model's reasoning in understandable terms to a human.

Most linear regression models, for example, are highly interpretable. (You merely need to look at the trained weights for each feature.) Decision forests are also highly interpretable. Some models, however, require sophisticated visualization to become interpretable.

You can use the Learning Interpretability Tool (LIT) to interpret ML models.

inter-rater agreement

A measurement of how often human raters agree when doing a task. If raters disagree, the task instructions may need to be improved. Also sometimes called inter-annotator agreement or inter-rater reliability . See also Cohen's kappa , which is one of the most popular inter-rater agreement measurements.

intersection over union (IoU)

#image

The intersection of two sets divided by their union. In machine-learning image-detection tasks, IoU is used to measure the accuracy of the model's predicted bounding box with respect to the ground-truth bounding box. In this case, the IoU for the two boxes is the ratio between the overlapping area and the total area, and its value ranges from 0 (no overlap of predicted bounding box and ground-truth bounding box) to 1 (predicted bounding box and ground-truth bounding box have the exact same coordinates).

For example, in the image below:

• The predicted bounding box (the coordinates delimiting where the model predicts the night table in the painting is located) is outlined in purple.
• The ground-truth bounding box (the coordinates delimiting where the night table in the painting is actually located) is outlined in green.

Here, the intersection of the bounding boxes for prediction and ground truth (below left) is 1, and the union of the bounding boxes for prediction and ground truth (below right) is 7, so the IoU is \(\frac{1}{7}\).

IoU

Abbreviation for intersection over union .

item matrix

#recsystems

In recommendation systems , a matrix of embedding vectors generated by matrix factorization that holds latent signals about each item . Each row of the item matrix holds the value of a single latent feature for all items. For example, consider a movie recommendation system. Each column in the item matrix represents a single movie. The latent signals might represent genres, or might be harder-to-interpret signals that involve complex interactions among genre, stars, movie age, or other factors.

The item matrix has the same number of columns as the target matrix that is being factorized. For example, given a movie recommendation system that evaluates 10,000 movie titles, the item matrix will have 10,000 columns.

موارد

#recsystems

In a recommendation system , the entities that a system recommends. For example, videos are the items that a video store recommends, while books are the items that a bookstore recommends.

تکرار

#fundamentals

A single update of a model's parameters—the model's weights and biases —during training . The batch size determines how many examples the model processes in a single iteration. For instance, if the batch size is 20, then the model processes 20 examples before adjusting the parameters.

When training a neural network , a single iteration involves the following two passes:

1. A forward pass to evaluate loss on a single batch.
2. A backward pass ( backpropagation ) to adjust the model's parameters based on the loss and the learning rate.

جی

JAX

An array computing library, bringing together XLA (Accelerated Linear Algebra) and automatic differentiation for high-performance numerical computing. JAX provides a simple and powerful API for writing accelerated numerical code with composable transformations. JAX provides features such as:

• grad (automatic differentiation)
• jit (just-in-time compilation)
• vmap (automatic vectorization or batching)
• pmap (parallelization)

JAX is a language for expressing and composing transformations of numerical code, analogous—but much larger in scope—to Python's NumPy library. (In fact, the .numpy library under JAX is a functionally equivalent, but entirely rewritten version of the Python NumPy library.)

JAX is particularly well-suited for speeding up many machine learning tasks by transforming the models and data into a form suitable for parallelism across GPU and TPU accelerator chips .

Flax , Optax , Pax , and many other libraries are built on the JAX infrastructure.

ک

کراس

A popular Python machine learning API. Keras runs on several deep learning frameworks, including TensorFlow, where it is made available as tf.keras .

Kernel Support Vector Machines (KSVMs)

A classification algorithm that seeks to maximize the margin between positive and negative classes by mapping input data vectors to a higher dimensional space. For example, consider a classification problem in which the input dataset has a hundred features. To maximize the margin between positive and negative classes, a KSVM could internally map those features into a million-dimension space. KSVMs uses a loss function called hinge loss .

امتیاز کلیدی

#image

The coordinates of particular features in an image. For example, for an image recognition model that distinguishes flower species, keypoints might be the center of each petal, the stem, the stamen, and so on.

k-fold cross validation

An algorithm for predicting a model's ability to generalize to new data. The k in k-fold refers to the number of equal groups you divide a dataset's examples into; that is, you train and test your model k times. For each round of training and testing, a different group is the test set, and all remaining groups become the training set. After k rounds of training and testing, you calculate the mean and standard deviation of the desired test metric(s).

For example, suppose your dataset consists of 120 examples. Further suppose, you decide to set k to 4. Therefore, after shuffling the examples, you divide the dataset into four equal groups of 30 examples and conduct four training/testing rounds:

For example, Mean Squared Error (MSE) might be the most meaningful metric for a linear regression model. Therefore, you would find the mean and standard deviation of the MSE across all four rounds.

k-به معنی

#clustering

A popular clustering algorithm that groups examples in unsupervised learning. The k-means algorithm basically does the following:

• Iteratively determines the best k center points (known as centroids ).
• Assigns each example to the closest centroid. Those examples nearest the same centroid belong to the same group.

The k-means algorithm picks centroid locations to minimize the cumulative square of the distances from each example to its closest centroid.

For example, consider the following plot of dog height to dog width:

If k=3, the k-means algorithm will determine three centroids. Each example is assigned to its closest centroid, yielding three groups:

Imagine that a manufacturer wants to determine the ideal sizes for small, medium, and large sweaters for dogs. The three centroids identify the mean height and mean width of each dog in that cluster. So, the manufacturer should probably base sweater sizes on those three centroids. Note that the centroid of a cluster is typically not an example in the cluster.

The preceding illustrations shows k-means for examples with only two features (height and width). Note that k-means can group examples across many features.

k-median

#clustering

A clustering algorithm closely related to k-means . The practical difference between the two is as follows:

• In k-means, centroids are determined by minimizing the sum of the squares of the distance between a centroid candidate and each of its examples.
• In k-median, centroids are determined by minimizing the sum of the distance between a centroid candidate and each of its examples.

Note that the definitions of distance are also different:

• k-means relies on the Euclidean distance from the centroid to an example. (In two dimensions, the Euclidean distance means using the Pythagorean theorem to calculate the hypotenuse.) For example, the k-means distance between (2,2) and (5,-2) would be:

$$ {\text{Euclidean distance}} = {\sqrt {(2-5)^2 + (2--2)^2}} = 5 $$

• k-median relies on the Manhattan distance from the centroid to an example. This distance is the sum of the absolute deltas in each dimension. For example, the k-median distance between (2,2) and (5,-2) would be:

$$ {\text{Manhattan distance}} = \lvert 2-5 \rvert + \lvert 2--2 \rvert = 7 $$

L

L ₀ regularization

#fundamentals

A type of regularization that penalizes the total number of nonzero weights in a model. For example, a model having 11 nonzero weights would be penalized more than a similar model having 10 nonzero weights.

L ₀ regularization is sometimes called L0-norm regularization .

Click the icon for additional notes.

L ₀ regularization is generally impractical in large models because L ₀ regularization turns training into a convex optimization problem.

---

L ₁ loss

#fundamentals

A loss function that calculates the absolute value of the difference between actual label values and the values that a model predicts. For example, here's the calculation of L ₁ loss for a batch of five examples :

Actual value of example	Model's predicted value	Absolute value of delta
7	6	1
5	4	1
8	11	3
4	6	2
9	8	1
		8 = L 1 loss

L ₁ loss is less sensitive to outliers than L ₂ loss .

The Mean Absolute Error is the average L ₁ loss per example.

Click the icon to see the formal math.

$$ L_1 loss = \sum_{i=0}^n | y_i - \hat{y}_i |$$

جایی که:

• $n$ is the number of examples.
• $y$ is the actual value of the label.
• $\hat{y}$ is the value that the model predicts for $y$.

---

L ₁ regularization

#fundamentals

A type of regularization that penalizes weights in proportion to the sum of the absolute value of the weights. L ₁ regularization helps drive the weights of irrelevant or barely relevant features to exactly 0 . A feature with a weight of 0 is effectively removed from the model.

Contrast with L ₂ regularization .

L ₂ loss

#fundamentals

A loss function that calculates the square of the difference between actual label values and the values that a model predicts. For example, here's the calculation of L ₂ loss for a batch of five examples :

Actual value of example	Model's predicted value	Square of delta
7	6	1
5	4	1
8	11	9
4	6	4
9	8	1
		16 = L 2 loss

Due to squaring, L ₂ loss amplifies the influence of outliers . That is, L ₂ loss reacts more strongly to bad predictions than L ₁ loss . For example, the L ₁ loss for the preceding batch would be 8 rather than 16. Notice that a single outlier accounts for 9 of the 16.

Regression models typically use L ₂ loss as the loss function.

The Mean Squared Error is the average L ₂ loss per example. Squared loss is another name for L ₂ loss.

Click the icon to see the formal math.

$$ L_2 loss = \sum_{i=0}^n {(y_i - \hat{y}_i)}^2$$

جایی که:

• $n$ is the number of examples.
• $y$ is the actual value of the label.
• $\hat{y}$ is the value that the model predicts for $y$.

---

L ₂ regularization

#fundamentals

A type of regularization that penalizes weights in proportion to the sum of the squares of the weights. L ₂ regularization helps drive outlier weights (those with high positive or low negative values) closer to 0 but not quite to 0 . Features with values very close to 0 remain in the model but don't influence the model's prediction very much.

L ₂ regularization always improves generalization in linear models .

Contrast with L ₁ regularization .

برچسب

#fundamentals

In supervised machine learning , the "answer" or "result" portion of an example .

Each labeled example consists of one or more features and a label. For instance, in a spam detection dataset, the label would probably be either "spam" or "not spam." In a rainfall dataset, the label might be the amount of rain that fell during a certain period.

labeled example

#fundamentals

An example that contains one or more features and a label . For example, the following table shows three labeled examples from a house valuation model, each with three features and one label:

تعداد اتاق خواب	تعداد حمام	House age	House price (label)
3	2	15	$345,000
2	1	72	$179,000
4	2	34	$392,000

In supervised machine learning , models train on labeled examples and make predictions on unlabeled examples .

Contrast labeled example with unlabeled examples.

label leakage

A model design flaw in which a feature is a proxy for the label . For example, consider a binary classification model that predicts whether or not a prospective customer will purchase a particular product. Suppose that one of the features for the model is a Boolean named SpokeToCustomerAgent . Further suppose that a customer agent is only assigned after the prospective customer has actually purchased the product. During training, the model will quickly learn the association between SpokeToCustomerAgent and the label.

لامبدا

#fundamentals

Synonym for regularization rate .

Lambda is an overloaded term. Here we're focusing on the term's definition within regularization .

LaMDA (Language Model for Dialogue Applications)

#زبان

A Transformer -based large language model developed by Google trained on a large dialogue dataset that can generate realistic conversational responses.

LaMDA: our breakthrough conversation technology provides an overview.

نقاط دیدنی

#image

Synonym for keypoints .

language model

#زبان

A model that estimates the probability of a token or sequence of tokens occurring in a longer sequence of tokens.

Click the icon for additional notes.

Though counterintuitive, many models that evaluate text are not language models . For example, text classification models and sentiment analysis models are not language models .

---

مدل زبان بزرگ

#زبان

An informal term with no strict definition that usually means a language model that has a high number of parameters . Some large language models contain over 100 billion parameters.

Click the icon for additional notes.

You might be wondering when a language model becomes large enough to be termed a large language model . Currently, there is no agreed-upon defining line for the number of parameters.

Most current large language models (for example, GPT ) are based on Transformer architecture.

---

لایه

#fundamentals

A set of neurons in a neural network . Three common types of layers are as follows:

• The input layer , which provides values for all the features .
• One or more hidden layers , which find nonlinear relationships between the features and the label.
• The output layer , which provides the prediction.

For example, the following illustration shows a neural network with one input layer, two hidden layers, and one output layer:

In TensorFlow , layers are also Python functions that take Tensors and configuration options as input and produce other tensors as output.

Layers API (tf.layers)

#TensorFlow

A TensorFlow API for constructing a deep neural network as a composition of layers. The Layers API enables you to build different types of layers , such as:

• tf.layers.Dense for a fully-connected layer .
• tf.layers.Conv2D for a convolutional layer.

The Layers API follows the Keras layers API conventions. That is, aside from a different prefix, all functions in the Layers API have the same names and signatures as their counterparts in the Keras layers API.

برگ

#df

Any endpoint in a decision tree . Unlike a condition , a leaf doesn't perform a test. Rather, a leaf is a possible prediction. A leaf is also the terminal node of an inference path .

For example, the following decision tree contains three leaves:

Learning Interpretability Tool (LIT)

A visual, interactive model-understanding and data visualization tool.

You can use open-source LIT to interpret models or to visualize text, image, and tabular data.

میزان یادگیری

#fundamentals

A floating-point number that tells the gradient descent algorithm how strongly to adjust weights and biases on each iteration . For example, a learning rate of 0.3 would adjust weights and biases three times more powerfully than a learning rate of 0.1.

Learning rate is a key hyperparameter . If you set the learning rate too low, training will take too long. If you set the learning rate too high, gradient descent often has trouble reaching convergence .

Click the icon for a more mathematical explanation.

During each iteration, the gradient descent algorithm multiplies the learning rate by the gradient. The resulting product is called the gradient step .

---

رگرسیون حداقل مربعات

A linear regression model trained by minimizing L ₂ Loss .

خطی

#fundamentals

A relationship between two or more variables that can be represented solely through addition and multiplication.

The plot of a linear relationship is a line.

Contrast with nonlinear .

مدل خطی

#fundamentals

A model that assigns one weight per feature to make predictions . (Linear models also incorporate a bias .) In contrast, the relationship of features to predictions in deep models is generally nonlinear .

Linear models are usually easier to train and more interpretable than deep models. However, deep models can learn complex relationships between features.

Linear regression and logistic regression are two types of linear models.

Click the icon to see the math.

A linear model follows this formula:

$$y' = b + w_1x_1 + w_2x_2 + … w_nx_n$$

جایی که:

• y' is the raw prediction. (In certain kinds of linear models, this raw prediction will be further modified. For example, see logistic regression .)
• b is the bias .
• w is a weight , so w ₁ is the weight of the first feature, w ₂ is the weight of the second feature, and so on.
• x is a feature , so x ₁ is the value of the first feature, x ₂ is the value of the second feature, and so on.

For example, suppose a linear model for three features learns the following bias and weights:

• b = 7
• w ₁ = -2.5
• w ₂ = -1.2
• w ₃ = 1.4

Therefore, given three features (x ₁ , x ₂ , and x ₃ ), the linear model uses the following equation to generate each prediction:

y' = 7 + (-2.5)(x1) + (-1.2)(x2) + (1.4)(x3)

Suppose a particular example contains the following values:

• x ₁ = 4
• x ₂ = -10
• x ₃ = 5

Plugging those values into the formula yields a prediction for this example:

y' = 7 + (-2.5)(4) + (-1.2)(-10) + (1.4)(5)
y' = 16

Linear models include not only models that use only a linear equation to make predictions but also a broader set of models that use a linear equation as just one component of the formula that makes predictions. For example, logistic regression post-processes the raw prediction (y') to produce a final prediction value between 0 and 1, exclusively.

---

رگرسیون خطی

#fundamentals

A type of machine learning model in which both of the following are true:

• The model is a linear model .
• The prediction is a floating-point value. (This is the regression part of linear regression .)

Contrast linear regression with logistic regression . Also, contrast regression with classification .

روشن

Abbreviation for the Learning Interpretability Tool (LIT) , which was previously known as the Language Interpretability Tool.

رگرسیون لجستیک

#fundamentals

A type of regression model that predicts a probability. Logistic regression models have the following characteristics:

• The label is categorical . The term logistic regression usually refers to binary logistic regression , that is, to a model that calculates probabilities for labels with two possible values. A less common variant, multinomial logistic regression , calculates probabilities for labels with more than two possible values.
• The loss function during training is Log Loss . (Multiple Log Loss units can be placed in parallel for labels with more than two possible values.)
• The model has a linear architecture, not a deep neural network. However, the remainder of this definition also applies to deep models that predict probabilities for categorical labels.

For example, consider a logistic regression model that calculates the probability of an input email being either spam or not spam. During inference, suppose the model predicts 0.72. Therefore, the model is estimating:

• A 72% chance of the email being spam.
• A 28% chance of the email not being spam.

A logistic regression model uses the following two-step architecture:

1. The model generates a raw prediction (y') by applying a linear function of input features.
2. The model uses that raw prediction as input to a sigmoid function , which converts the raw prediction to a value between 0 and 1, exclusive.

Like any regression model, a logistic regression model predicts a number. However, this number typically becomes part of a binary classification model as follows:

• If the predicted number is greater than the classification threshold , the binary classification model predicts the positive class.
• If the predicted number is less than the classification threshold, the binary classification model predicts the negative class.

logits

The vector of raw (non-normalized) predictions that a classification model generates, which is ordinarily then passed to a normalization function. If the model is solving a multi-class classification problem, logits typically become an input to the softmax function. The softmax function then generates a vector of (normalized) probabilities with one value for each possible class.

tf.nn.sigmoid_cross_entropy_with_logits .

Log Loss

#fundamentals

The loss function used in binary logistic regression .

Click the icon to see the math.

The following formula calculates Log Loss:

$$\text{Log Loss} = \sum_{(x,y)\in D} -y\log(y') - (1 - y)\log(1 - y')$$

جایی که:

• \((x,y)\in D\) is the data set containing many labeled examples, which are \((x,y)\) pairs.
• \(y\) is the label in a labeled example. Since this is logistic regression, every value of \(y\) must either be 0 or 1.
• \(y'\) is the predicted value (somewhere between 0 and 1, exclusive), given the set of features in \(x\).

---

log-odds

#fundamentals

The logarithm of the odds of some event.

Click the icon to see the math.

If the event is a binary probability, then odds refers to the ratio of the probability of success ( p ) to the probability of failure (1- p ). For example, suppose that a given event has a 90% probability of success and a 10% probability of failure. In this case, odds is calculated as follows:

$$ {\text{odds}} = \frac{\text{p}} {\text{(1-p)}} = \frac{.9} {.1} = {\text{9}} $$

The log-odds is simply the logarithm of the odds. By convention, "logarithm" refers to natural logarithm , but logarithm could actually be any base greater than 1. Sticking to convention, the log-odds of our example is therefore:

$$ {\text{log-odds}} = ln(9) ~= 2.2 $$

The log-odds function is the inverse of the sigmoid function .

---

Long Short-Term Memory (LSTM)

#seq

A type of cell in a recurrent neural network used to process sequences of data in applications such as handwriting recognition, machine translation, and image captioning. LSTMs address the vanishing gradient problem that occurs when training RNNs due to long data sequences by maintaining history in an internal memory state based on new input and context from previous cells in the RNN.

ضرر - زیان

#fundamentals

During the training of a supervised model , a measure of how far a model's prediction is from its label .

A loss function calculates the loss.

loss aggregator

A type of machine learning algorithm that improves the performance of a model by combining the predictions of multiple models and using those predictions to make a single prediction. As a result, a loss aggregator can reduce the variance of the predictions and improve the accuracy of the predictions.

loss curve

#fundamentals

A plot of loss as a function of the number of training iterations . The following plot shows a typical loss curve:

Loss curves can help you determine when your model is converging or overfitting .

Loss curves can plot all of the following types of loss:

• training loss
• validation loss
• test loss

See also generalization curve .

loss function

#fundamentals

During training or testing, a mathematical function that calculates the loss on a batch of examples. A loss function returns a lower loss for models that makes good predictions than for models that make bad predictions.

The goal of training is typically to minimize the loss that a loss function returns.

Many different kinds of loss functions exist. Pick the appropriate loss function for the kind of model you are building. مثلا:

• L ₂ loss (or Mean Squared Error ) is the loss function for linear regression .
• Log Loss is the loss function for logistic regression .

loss surface

A graph of weight(s) vs. loss. Gradient descent aims to find the weight(s) for which the loss surface is at a local minimum.

LSTM

#seq

Abbreviation for Long Short-Term Memory .

م

فراگیری ماشین

#fundamentals

A program or system that trains a model from input data. The trained model can make useful predictions from new (never-before-seen) data drawn from the same distribution as the one used to train the model.

Machine learning also refers to the field of study concerned with these programs or systems.

majority class

#fundamentals

The more common label in a class-imbalanced dataset . For example, given a dataset containing 99% negative labels and 1% positive labels, the negative labels are the majority class.

Contrast with minority class .

Markov decision process (MDP)

#rl

A graph representing the decision-making model where decisions (or actions ) are taken to navigate a sequence of states under the assumption that the Markov property holds. In reinforcement learning , these transitions between states return a numerical reward .

دارایی مارکوف

#rl

A property of certain environments , where state transitions are entirely determined by information implicit in the current state and the agent's action .

masked language model

#زبان

A language model that predicts the probability of candidate tokens to fill in blanks in a sequence. For instance, a masked language model can calculate probabilities for candidate word(s) to replace the underline in the following sentence:

The ____ in the hat came back.

The literature typically uses the string "MASK" instead of an underline. مثلا:

The "MASK" in the hat came back.

Most modern masked language models are bidirectional .

matplotlib

An open-source Python 2D plotting library. matplotlib helps you visualize different aspects of machine learning.

فاکتورسازی ماتریسی

#recsystems

In math, a mechanism for finding the matrices whose dot product approximates a target matrix.

In recommendation systems , the target matrix often holds users' ratings on items . For example, the target matrix for a movie recommendation system might look something like the following, where the positive integers are user ratings and 0 means that the user didn't rate the movie:

	کازابلانکا	داستان فیلادلفیا	پلنگ سیاه	زن شگفت انگیز	داستان عامهپسند
کاربر 1	5.0	3.0	0.0	2.0	0.0
User 2	4.0	0.0	0.0	1.0	5.0
کاربر 3	3.0	1.0	4.0	5.0	0.0

The movie recommendation system aims to predict user ratings for unrated movies. For example, will User 1 like Black Panther ?

One approach for recommendation systems is to use matrix factorization to generate the following two matrices:

• A user matrix , shaped as the number of users X the number of embedding dimensions.
• An item matrix , shaped as the number of embedding dimensions X the number of items.

For example, using matrix factorization on our three users and five items could yield the following user matrix and item matrix:

User Matrix                 Item Matrix

1.1   2.3           0.9   0.2   1.4    2.0   1.2
0.6   2.0           1.7   1.2   1.2   -0.1   2.1
2.5   0.5

The dot product of the user matrix and item matrix yields a recommendation matrix that contains not only the original user ratings but also predictions for the movies that each user hasn't seen. For example, consider User 1's rating of Casablanca , which was 5.0. The dot product corresponding to that cell in the recommendation matrix should hopefully be around 5.0, and it is:

(1.1 * 0.9) + (2.3 * 1.7) = 4.9

More importantly, will User 1 like Black Panther ? Taking the dot product corresponding to the first row and the third column yields a predicted rating of 4.3:

(1.1 * 1.4) + (2.3 * 1.2) = 4.3

Matrix factorization typically yields a user matrix and item matrix that, together, are significantly more compact than the target matrix.

میانگین خطای مطلق (MAE)

The average loss per example when L ₁ loss is used. Calculate Mean Absolute Error as follows:

1. Calculate the L ₁ loss for a batch.
2. Divide the L ₁ loss by the number of examples in the batch.

Click the icon to see the formal math.

$$\text{Mean Absolute Error} = \frac{1}{n}\sum_{i=0}^n | y_i - \hat{y}_i |$$where:

• $n$ is the number of examples.
• $y$ is the actual value of the label.
• $\hat{y}$ is the value that the model predicts for $y$.

---

For example, consider the calculation of L ₁ loss on the following batch of five examples:

Actual value of example	Model's predicted value	Loss (difference between actual and predicted)
7	6	1
5	4	1
8	11	3
4	6	2
9	8	1
		8 = L 1 loss

So, L ₁ loss is 8 and the number of examples is 5. Therefore, the Mean Absolute Error is:

Mean Absolute Error = L1 loss / Number of Examples
Mean Absolute Error = 8/5 = 1.6

Contrast Mean Absolute Error with Mean Squared Error and Root Mean Squared Error .

میانگین مربعات خطا (MSE)

The average loss per example when L ₂ loss is used. Calculate Mean Squared Error as follows:

1. Calculate the L ₂ loss for a batch.
2. Divide the L ₂ loss by the number of examples in the batch.

Click the icon to see the formal math.

$$\text{Mean Squared Error} = \frac{1}{n}\sum_{i=0}^n {(y_i - \hat{y}_i)}^2$$where:

• $n$ is the number of examples.
• $y$ is the actual value of the label.
• $\hat{y}$ is the model's prediction for $y$.

---

For example, consider the loss on the following batch of five examples:

ارزش واقعی	Model's prediction	ضرر - زیان	Squared loss
7	6	1	1
5	4	1	1
8	11	3	9
4	6	2	4
9	8	1	1
			16 = L 2 loss

Therefore, the Mean Squared Error is:

Mean Squared Error = L2 loss / Number of Examples
Mean Squared Error = 16/5 = 3.2

Mean Squared Error is a popular training optimizer , particularly for linear regression .

Contrast Mean Squared Error with Mean Absolute Error and Root Mean Squared Error .

TensorFlow Playground uses Mean Squared Error to calculate loss values.

Click the icon to see more details about outliers.

Outliers strongly influence Mean Squared Error. For example, a loss of 1 is a squared loss of 1, but a loss of 3 is a squared loss of 9. In the preceding table, the example with a loss of 3 accounts for ~56% of the Mean Squared Error, while each of the examples with a loss of 1 accounts for only 6% of the Mean Squared Error.

Outliers do not influence Mean Absolute Error as strongly as Mean Squared Error. For example, a loss of 3 accounts for only ~38% of the Mean Absolute Error.

Clipping is one way to prevent extreme outliers from damaging your model's predictive ability.

---

مش

#TensorFlow

#GoogleCloud

In ML parallel programming, a term associated with assigning the data and model to TPU chips, and defining how these values will be sharded or replicated.

Mesh is an overloaded term that can mean either of the following:

• A physical layout of TPU chips.
• An abstract logical construct for mapping the data and model to the TPU chips.

In either case, a mesh is specified as a shape .

meta-learning

#زبان

A subset of machine learning that discovers or improves a learning algorithm. A meta-learning system can also aim to train a model to quickly learn a new task from a small amount of data or from experience gained in previous tasks. Meta-learning algorithms generally try to achieve the following:

• Improve/learn hand-engineered features (such as an initializer or an optimizer).
• Be more data-efficient and compute-efficient.
• Improve generalization.

Meta-learning is related to few-shot learning .

متریک

#TensorFlow

A statistic that you care about.

An objective is a metric that a machine learning system tries to optimize.

Metrics API (tf.metrics)

A TensorFlow API for evaluating models. For example, tf.metrics.accuracy determines how often a model's predictions match labels.

mini-batch

#fundamentals

A small, randomly selected subset of a batch processed in one iteration . The batch size of a mini-batch is usually between 10 and 1,000 examples.

For example, suppose the entire training set (the full batch) consists of 1,000 examples. Further suppose that you set the batch size of each mini-batch to 20. Therefore, each iteration determines the loss on a random 20 of the 1,000 examples and then adjusts the weights and biases accordingly.

It is much more efficient to calculate the loss on a mini-batch than the loss on all the examples in the full batch.

mini-batch stochastic gradient descent

A gradient descent algorithm that uses mini-batches . In other words, mini-batch stochastic gradient descent estimates the gradient based on a small subset of the training data. Regular stochastic gradient descent uses a mini-batch of size 1.

minimax loss

A loss function for generative adversarial networks , based on the cross-entropy between the distribution of generated data and real data.

Minimax loss is used in the first paper to describe generative adversarial networks.

minority class

#fundamentals

The less common label in a class-imbalanced dataset . For example, given a dataset containing 99% negative labels and 1% positive labels, the positive labels are the minority class.

Contrast with majority class .

Click the icon for additional notes.

A training set with a million examples sounds impressive. However, if the minority class is poorly represented, then even a very large training set might be insufficient. Focus less on the total number of examples in the dataset and more on the number of examples in the minority class.

If your dataset doesn't contain enough minority class examples, consider using downsampling (the definition in the second bullet) to supplement the minority class.

---

ML

Abbreviation for machine learning .

MNIST

#image

A public-domain dataset compiled by LeCun, Cortes, and Burges containing 60,000 images, each image showing how a human manually wrote a particular digit from 0–9. Each image is stored as a 28x28 array of integers, where each integer is a grayscale value between 0 and 255, inclusive.

MNIST is a canonical dataset for machine learning, often used to test new machine learning approaches. For details, see The MNIST Database of Handwritten Digits .

روش

#زبان

A high-level data category. For example, numbers, text, images, video, and audio are five different modalities.

مدل

#fundamentals

In general, any mathematical construct that processes input data and returns output. Phrased differently, a model is the set of parameters and structure needed for a system to make predictions. In supervised machine learning , a model takes an example as input and infers a prediction as output. Within supervised machine learning, models differ somewhat. مثلا:

• A linear regression model consists of a set of weights and a bias .
• A neural network model consists of:
  • A set of hidden layers , each containing one or more neurons .
  • The weights and bias associated with each neuron.
• A decision tree model consists of:
  • The shape of the tree; that is, the pattern in which the conditions and leaves are connected.
  • The conditions and leaves.

You can save, restore, or make copies of a model.

Unsupervised machine learning also generates models, typically a function that can map an input example to the most appropriate cluster .

Click the icon to compare algebraic and programming functions to ML models.

An algebraic function such as the following is a model:

  f(x, y) = 3x -5xy + y2 + 17

The preceding function maps input values ( x and y ) to output.

Similarly, a programming function like the following is also a model:

def half_of_greater(x, y):
  if (x > y):
    return(x / 2)
  else
    return(y / 2)

A caller passes arguments to the preceding Python function, and the Python function generates output (via the return statement).

Although a deep neural network has a very different mathematical structure than an algebraic or programming function, a deep neural network still takes input (an example) and returns output (a prediction).

A human programmer codes a programming function manually. In contrast, a machine learning model gradually learns the optimal parameters during automated training.

---

model capacity

The complexity of problems that a model can learn. The more complex the problems that a model can learn, the higher the model's capacity. A model's capacity typically increases with the number of model parameters. For a formal definition of classifier capacity, see VC dimension .

model parallelism

#زبان

A way of scaling training or inference that puts different parts of one model on different devices . Model parallelism enables models that are too big to fit on a single device.

To implement model parallelism, a system typically does the following:

1. Shards (divides) the model into smaller parts.
2. Distributes the training of those smaller parts across multiple processors. Each processor trains its own part of the model.
3. Combines the results to create a single model.

Model parallelism slows training.

See also data parallelism .

آموزش مدل

The process of determining the best model .

تکانه

A sophisticated gradient descent algorithm in which a learning step depends not only on the derivative in the current step, but also on the derivatives of the step(s) that immediately preceded it. Momentum involves computing an exponentially weighted moving average of the gradients over time, analogous to momentum in physics. Momentum sometimes prevents learning from getting stuck in local minima.

multi-class classification

#fundamentals

In supervised learning, a classification problem in which the dataset contains more than two classes of labels. For example, the labels in the Iris dataset must be one of the following three classes:

• Iris setosa
• Iris virginica
• Iris versicolor

A model trained on the Iris dataset that predicts Iris type on new examples is performing multi-class classification.

In contrast, classification problems that distinguish between exactly two classes are binary classification models . For example, an email model that predicts either spam or not spam is a binary classification model.

In clustering problems, multi-class classification refers to more than two clusters.

multi-class logistic regression

Using logistic regression in multi-class classification problems.

multi-head self-attention

#زبان

An extension of self-attention that applies the self-attention mechanism multiple times for each position in the input sequence.

Transformers introduced multi-head self-attention.

multimodal model

#زبان

A model whose inputs and/or outputs include more than one modality . For example, consider a model that takes both an image and a text caption (two modalities) as features , and outputs a score indicating how appropriate the text caption is for the image. So, this model's inputs are multimodal and the output is unimodal.

multinomial classification

Synonym for multi-class classification .

multinomial regression

Synonym for multi-class logistic regression .

چند وظیفه ای

A machine learning technique in which a single model is trained to perform multiple tasks .

Multitask models are created by training on data that is appropriate for each of the different tasks. This allows the model to learn to share information across the tasks, which helps the model learn more effectively.

A model trained for multiple tasks often has improved generalization abilities and can be more robust at handling different types of data.

ن

NaN trap

When one number in your model becomes a NaN during training, which causes many or all other numbers in your model to eventually become a NaN.

NaN is an abbreviation for N ot a N umber.

natural language understanding

#زبان

Determining a user's intentions based on what the user typed or said. For example, a search engine uses natural language understanding to determine what the user is searching for based on what the user typed or said.

negative class

#fundamentals

In binary classification , one class is termed positive and the other is termed negative . The positive class is the thing or event that the model is testing for and the negative class is the other possibility. مثلا:

• The negative class in a medical test might be "not tumor."
• The negative class in an email classifier might be "not spam."

Contrast with positive class .

negative sampling

Synonym for candidate sampling .

Neural Architecture Search (NAS)

A technique for automatically designing the architecture of a neural network . NAS algorithms can reduce the amount of time and resources required to train a neural network.

NAS typically uses:

• A search space, which is a set of possible architectures.
• A fitness function, which is a measure of how well a particular architecture performs on a given task.

NAS algorithms often start with a small set of possible architectures and gradually expand the search space as the algorithm learns more about what architectures are effective. The fitness function is typically based on the performance of the architecture on a training set, and the algorithm is typically trained using a reinforcement learning technique.

NAS algorithms have proven effective in finding high-performing architectures for a variety of tasks, including image classification , text classification, and machine translation.

شبکه عصبی

#fundamentals

A model containing at least one hidden layer . A deep neural network is a type of neural network containing more than one hidden layer. For example, the following diagram shows a deep neural network containing two hidden layers.

Each neuron in a neural network connects to all of the nodes in the next layer. For example, in the preceding diagram, notice that each of the three neurons in the first hidden layer separately connect to both of the two neurons in the second hidden layer.

Neural networks implemented on computers are sometimes called artificial neural networks to differentiate them from neural networks found in brains and other nervous systems.

Some neural networks can mimic extremely complex nonlinear relationships between different features and the label.

See also convolutional neural network and recurrent neural network .

نورون

#fundamentals

In machine learning, a distinct unit within a hidden layer of a neural network . Each neuron performs the following two-step action:

1. Calculates the weighted sum of input values multiplied by their corresponding weights.
2. Passes the weighted sum as input to an activation function .

A neuron in the first hidden layer accepts inputs from the feature values in the input layer . A neuron in any hidden layer beyond the first accepts inputs from the neurons in the preceding hidden layer. For example, a neuron in the second hidden layer accepts inputs from the neurons in the first hidden layer.

The following illustration highlights two neurons and their inputs.

A neuron in a neural network mimics the behavior of neurons in brains and other parts of nervous systems.

N-گرم

#seq

#زبان

An ordered sequence of N words. For example, truly madly is a 2-gram. Because order is relevant, madly truly is a different 2-gram than truly madly .

ن	Name(s) for this kind of N-gram	مثال ها
2	bigram or 2-gram	to go, go to, eat lunch, eat dinner
3	trigram or 3-gram	ate too much, three blind mice, the bell tolls
4	4-gram	walk in the park, dust in the wind, the boy ate lentils

Many natural language understanding models rely on N-grams to predict the next word that the user will type or say. For example, suppose a user typed three blind . An NLU model based on trigrams would likely predict that the user will next type mice .

Contrast N-grams with bag of words , which are unordered sets of words.

NLU

#زبان

Abbreviation for natural language understanding .

node (decision tree)

#df

In a decision tree , any condition or leaf .

node (neural network)

#fundamentals

A neuron in a hidden layer .

node (TensorFlow graph)

#TensorFlow

An operation in a TensorFlow graph .

سر و صدا

Broadly speaking, anything that obscures the signal in a dataset. Noise can be introduced into data in a variety of ways. مثلا:

• Human raters make mistakes in labeling.
• Humans and instruments mis-record or omit feature values.

non-binary condition

#df

A condition containing more than two possible outcomes. For example, the following non-binary condition contains three possible outcomes:

غیر خطی

#fundamentals

A relationship between two or more variables that can't be represented solely through addition and multiplication. A linear relationship can be represented as a line; a nonlinear relationship can't be represented as a line. For example, consider two models that each relate a single feature to a single label. The model on the left is linear and the model on the right is nonlinear:

non-response bias

#fairness

See selection bias .

نامانایی

#fundamentals

A feature whose values change across one or more dimensions, usually time. For example, consider the following examples of nonstationarity:

• The number of swimsuits sold at a particular store varies with the season.
• The quantity of a particular fruit harvested in a particular region is zero for much of the year but large for a brief period.
• Due to climate change, annual mean temperatures are shifting.

Contrast with stationarity .

عادی سازی

#fundamentals

Broadly speaking, the process of converting a variable's actual range of values into a standard range of values, such as:

• -1 to +1
• 0 to 1
• the normal distribution

For example, suppose the actual range of values of a certain feature is 800 to 2,400. As part of feature engineering , you could normalize the actual values down to a standard range, such as -1 to +1.

Normalization is a common task in feature engineering . Models usually train faster (and produce better predictions) when every numerical feature in the feature vector has roughly the same range.

novelty detection

The process of determining whether a new (novel) example comes from the same distribution as the training set . In other words, after training on the training set, novelty detection determines whether a new example (during inference or during additional training) is an outlier .

Contrast with outlier detection .

داده های عددی

#fundamentals

Features represented as integers or real-valued numbers. For example, a house valuation model would probably represent the size of a house (in square feet or square meters) as numerical data. Representing a feature as numerical data indicates that the feature's values have a mathematical relationship to the label. That is, the number of square meters in a house probably has some mathematical relationship to the value of the house.

Not all integer data should be represented as numerical data. For example, postal codes in some parts of the world are integers; however, integer postal codes should not be represented as numerical data in models. That's because a postal code of 20000 is not twice (or half) as potent as a postal code of 10000. Furthermore, although different postal codes do correlate to different real estate values, we can't assume that real estate values at postal code 20000 are twice as valuable as real estate values at postal code 10000. Postal codes should be represented as categorical data instead.

Numerical features are sometimes called continuous features .

NumPy

An open-source math library that provides efficient array operations in Python. pandas is built on NumPy.

O

هدف، واقعگرایانه

A metric that your algorithm is trying to optimize.

تابع هدف

The mathematical formula or metric that a model aims to optimize. For example, the objective function for linear regression is usually Mean Squared Loss . Therefore, when training a linear regression model, training aims to minimize Mean Squared Loss.

In some cases, the goal is to maximize the objective function. For example, if the objective function is accuracy, the goal is to maximize accuracy.

See also loss .

oblique condition

#df

In a decision tree , a condition that involves more than one feature . For example, if height and width are both features, then the following is an oblique condition:

  height > width

Contrast with axis-aligned condition .

آفلاین

#fundamentals

Synonym for static .

offline inference

#fundamentals

The process of a model generating a batch of predictions and then caching (saving) those predictions. Apps can then access the desired prediction from the cache rather than rerunning the model.

For example, consider a model that generates local weather forecasts (predictions) once every four hours. After each model run, the system caches all the local weather forecasts. Weather apps retrieve the forecasts from the cache.

Offline inference is also called static inference .

Contrast with online inference .

رمزگذاری تک داغ

#fundamentals

Representing categorical data as a vector in which:

• One element is set to 1.
• All other elements are set to 0.

One-hot encoding is commonly used to represent strings or identifiers that have a finite set of possible values. For example, suppose a certain categorical feature named Scandinavia has five possible values:

• "دانمارک"
• "سوئد"
• "Norway"
• "Finland"
• "ایسلند"

One-hot encoding could represent each of the five values as follows:

کشور	بردار
"دانمارک"	1	0	0	0	0
"سوئد"	0	1	0	0	0
"Norway"	0	0	1	0	0
"Finland"	0	0	0	1	0
"ایسلند"	0	0	0	0	1

Thanks to one-hot encoding, a model can learn different connections based on each of the five countries.

Representing a feature as numerical data is an alternative to one-hot encoding. Unfortunately, representing the Scandinavian countries numerically is not a good choice. For example, consider the following numeric representation:

• "Denmark" is 0
• "Sweden" is 1
• "Norway" is 2
• "Finland" is 3
• "Iceland" is 4

With numeric encoding, a model would interpret the raw numbers mathematically and would try to train on those numbers. However, Iceland isn't actually twice as much (or half as much) of something as Norway, so the model would come to some strange conclusions.

one-shot learning

A machine learning approach, often used for object classification, designed to learn effective classifiers from a single training example.

See also few-shot learning and zero-shot learning .

one-shot prompting

#زبان

#generativeAI

A prompt that contains one example demonstrating how the large language model should respond. For example, the following prompt contains one example showing a large language model how it should answer a query.

Parts of one prompt	یادداشت
What is the official currency of the specified country?	The question you want the LLM to answer.
France: EUR	یک مثال.
هند:	The actual query.

Compare and contrast one-shot prompting with the following terms:

• zero-shot prompting
• few-shot prompting

one-vs.-all

#fundamentals

Given a classification problem with N classes, a solution consisting of N separate binary classifiers —one binary classifier for each possible outcome. For example, given a model that classifies examples as animal, vegetable, or mineral, a one-vs.-all solution would provide the following three separate binary classifiers:

• animal vs. not animal
• vegetable vs. not vegetable
• mineral vs. not mineral

برخط

#fundamentals

Synonym for dynamic .

online inference

#fundamentals

Generating predictions on demand. For example, suppose an app passes input to a model and issues a request for a prediction. A system using online inference responds to the request by running the model (and returning the prediction to the app).

Contrast with offline inference .

operation (op)

#TensorFlow

In TensorFlow, any procedure that creates, manipulates, or destroys a Tensor . For example, a matrix multiply is an operation that takes two Tensors as input and generates one Tensor as output.

Optax

A gradient processing and optimization library for JAX . Optax facilitates research by providing building blocks that can be recombined in custom ways to optimize parametric models such as deep neural networks. اهداف دیگر عبارتند از:

• Providing readable, well-tested, efficient implementations of core components.
• Improving productivity by making it possible to combine low level ingredients into custom optimizers (or other gradient processing components).
• Accelerating adoption of new ideas by making it easy for anyone to contribute.

بهینه ساز

A specific implementation of the gradient descent algorithm. Popular optimizers include:

• AdaGrad , which stands for ADAptive GRADient descent.
• Adam, which stands for ADAptive with Momentum.

out-group homogeneity bias

#fairness

The tendency to see out-group members as more alike than in-group members when comparing attitudes, values, personality traits, and other characteristics. In-group refers to people you interact with regularly; out-group refers to people you do not interact with regularly. If you create a dataset by asking people to provide attributes about out-groups, those attributes may be less nuanced and more stereotyped than attributes that participants list for people in their in-group.

For example, Lilliputians might describe the houses of other Lilliputians in great detail, citing small differences in architectural styles, windows, doors, and sizes. However, the same Lilliputians might simply declare that Brobdingnagians all live in identical houses.

Out-group homogeneity bias is a form of group attribution bias .

See also in-group bias .

تشخیص بیرونی

The process of identifying outliers in a training set .

Contrast with novelty detection .

موارد پرت

Values distant from most other values. In machine learning, any of the following are outliers:

• Input data whose values are more than roughly 3 standard deviations from the mean.
• Weights with high absolute values.
• Predicted values relatively far away from the actual values.

For example, suppose that widget-price is a feature of a certain model. Assume that the mean widget-price is 7 Euros with a standard deviation of 1 Euro. Examples containing a widget-price of 12 Euros or 2 Euros would therefore be considered outliers because each of those prices is five standard deviations from the mean.

Outliers are often caused by typos or other input mistakes. In other cases, outliers aren't mistakes; after all, values five standard deviations away from the mean are rare but hardly impossible.

Outliers often cause problems in model training. Clipping is one way of managing outliers.

out-of-bag evaluation (OOB evaluation)

#df

A mechanism for evaluating the quality of a decision forest by testing each decision tree against the examples not used during training of that decision tree. For example, in the following diagram, notice that the system trains each decision tree on about two-thirds of the examples and then evaluates against the remaining one-third of the examples.

Out-of-bag evaluation is a computationally efficient and conservative approximation of the cross-validation mechanism. In cross-validation, one model is trained for each cross-validation round (for example, 10 models are trained in a 10-fold cross-validation). With OOB evaluation, a single model is trained. Because bagging withholds some data from each tree during training, OOB evaluation can use that data to approximate cross-validation.

output layer

#fundamentals

The "final" layer of a neural network. The output layer contains the prediction.

The following illustration shows a small deep neural network with an input layer, two hidden layers, and an output layer:

بیش از حد

#fundamentals

Creating a model that matches the training data so closely that the model fails to make correct predictions on new data.

Regularization can reduce overfitting. Training on a large and diverse training set can also reduce overfitting.

Click the icon for additional notes.

Overfitting is like strictly following advice from only your favorite teacher. You'll probably be successful in that teacher's class, but you might "overfit" to that teacher's ideas and be unsuccessful in other classes. Following advice from a mixture of teachers will enable you to adapt better to new situations.

---

oversampling

Reusing the examples of a minority class in a class-imbalanced dataset in order to create a more balanced training set .

For example, consider a binary classification problem in which the ratio of the majority class to the minority class is 5,000:1. If the dataset contains a million examples, then the dataset contains only about 200 examples of the minority class, which might be too few examples for effective training. To overcome this deficiency, you might oversample (reuse) those 200 examples multiple times, possibly yielding sufficient examples for useful training.

You need to be careful about over overfitting when oversampling.

Contrast with undersampling .

پ

packed data

An approach for storing data more efficiently.

Packed data stores data either by using a compressed format or in some other way that allows it to be accessed more efficiently. Packed data minimizes the amount of memory and computation required to access it, leading to faster training and more efficient model inference.

Packed data is often used with other techniques, such as data augmentation and regularization , further improving the performance of models .

پانداها

#fundamentals

A column-oriented data analysis API built on top of numpy . Many machine learning frameworks, including TensorFlow, support pandas data structures as inputs. See the pandas documentation for details.

پارامتر

#fundamentals

The weights and biases that a model learns during training . For example, in a linear regression model, the parameters consist of the bias ( b ) and all the weights ( w ₁ , w ₂ , and so on) in the following formula:

$$y' = b + w_1x_1 + w_2x_2 + … w_nx_n$$

In contrast, hyperparameter are the values that you (or a hyperparameter turning service) supply to the model. For example, learning rate is a hyperparameter.

parameter-efficient tuning

#زبان

#generativeAI

A set of techniques to fine-tune a large pre-trained language model (PLM) more efficiently than full fine-tuning . Parameter-efficient tuning typically fine-tunes far fewer parameters than full fine-tuning, yet generally produces a large language model that performs as well (or almost as well) as a large language model built from full fine-tuning.

Compare and contrast parameter-efficient tuning with:

• instruction tuning
• prompt tuning

Parameter-efficient tuning is also known as parameter-efficient fine-tuning .

Parameter Server (PS)

#TensorFlow

A job that keeps track of a model's parameters in a distributed setting.

parameter update

The operation of adjusting a model's parameters during training, typically within a single iteration of gradient descent .

partial derivative

A derivative in which all but one of the variables is considered a constant. For example, the partial derivative of f(x, y) with respect to x is the derivative of f considered as a function of x alone (that is, keeping y constant). The partial derivative of f with respect to x focuses only on how x is changing and ignores all other variables in the equation.

participation bias

#fairness

Synonym for non-response bias. See selection bias .

partitioning strategy

The algorithm by which variables are divided across parameter servers .

پکس

A programming framework designed for training large-scale neural network models so large that they span multiple TPU accelerator chip slices or pods .

Pax is built on Flax , which is built on JAX .

پرسپترون

A system (either hardware or software) that takes in one or more input values, runs a function on the weighted sum of the inputs, and computes a single output value. In machine learning, the function is typically nonlinear, such as ReLU , sigmoid , or tanh . For example, the following perceptron relies on the sigmoid function to process three input values:

$$f(x_1, x_2, x_3) = \text{sigmoid}(w_1 x_1 + w_2 x_2 + w_3 x_3)$$

In the following illustration, the perceptron takes three inputs, each of which is itself modified by a weight before entering the perceptron:

Perceptrons are the neurons in neural networks .

کارایی

Overloaded term with the following meanings:

• The traditional meaning within software engineering. Namely: How fast (or efficiently) does this piece of software run?
• The meaning within machine learning. Here, performance answers the following question: How correct is this model ? That is, how good are the model's predictions?

permutation variable importances

#df

A type of variable importance that evaluates the increase in the prediction error of a model after permuting the feature's values. Permutation variable importance is a model agnostic metric.

گیجی

One measure of how well a model is accomplishing its task. For example, suppose your task is to read the first few letters of a word a user is typing on a smartphone keyboard, and to offer a list of possible completion words. Perplexity, P, for this task is approximately the number of guesses you need to offer in order for your list to contain the actual word the user is trying to type.

Perplexity is related to cross-entropy as follows:

$$P= 2^{-\text{cross entropy}}$$

خط لوله

The infrastructure surrounding a machine learning algorithm. A pipeline includes gathering the data, putting the data into training data files, training one or more models, and exporting the models to production.

خط لوله

#زبان

A form of model parallelism in which a model's processing is divided into consecutive stages and each stage is executed on a different device. While a stage is processing one batch, the preceding stage can work on the next batch.

See also staged training .

pjit

A JAX function that splits code to run across multiple accelerator chips . The user passes a function to pjit, which returns a function that has the equivalent semantics but is compiled into an XLA computation that runs across multiple devices (such as GPUs or TPU cores).

pjit enables users to shard computations without rewriting them by using the SPMD partitioner.

As of March 2023, pjit has been merged with jit . Refer to Distributed arrays and automatic parallelization for more details.

PLM

#زبان

#generativeAI

Abbreviation for pre-trained language model .

pmap

A JAX function that executes copies of an input function on multiple underlying hardware devices (CPUs, GPUs, or TPUs ), with different input values. pmap relies on SPMD .

خط مشی

#rl

In reinforcement learning, an agent's probabilistic mapping from states to actions .

ادغام

#image

Reducing a matrix (or matrices) created by an earlier convolutional layer to a smaller matrix. Pooling usually involves taking either the maximum or average value across the pooled area. For example, suppose we have the following 3x3 matrix:

A pooling operation, just like a convolutional operation, divides that matrix into slices and then slides that convolutional operation by strides . For example, suppose the pooling operation divides the convolutional matrix into 2x2 slices with a 1x1 stride. As the following diagram illustrates, four pooling operations take place. Imagine that each pooling operation picks the maximum value of the four in that slice:

Pooling helps enforce translational invariance in the input matrix.

Pooling for vision applications is known more formally as spatial pooling . Time-series applications usually refer to pooling as temporal pooling . Less formally, pooling is often called subsampling or downsampling .

positional encoding

#زبان

A technique to add information about the position of a token in a sequence to the token's embedding. Transformer models use positional encoding to better understand the relationship between different parts of the sequence.

A common implementation of positional encoding uses a sinusoidal function. (Specifically, the frequency and amplitude of the sinusoidal function are determined by the position of the token in the sequence.) This technique enables a Transformer model to learn to attend to different parts of the sequence based on their position.

positive class

#fundamentals

The class you are testing for.

For example, the positive class in a cancer model might be "tumor." The positive class in an email classifier might be "spam."

Contrast with negative class .

Click the icon for additional notes.

The term positive class can be confusing because the "positive" outcome of many tests is often an undesirable result. For example, the positive class in many medical tests corresponds to tumors or diseases. In general, you want a doctor to tell you, "Congratulations! Your test results were negative." Regardless, the positive class is the event that the test is seeking to find.

Admittedly, you're simultaneously testing for both the positive and negative classes.

---

پس پردازش

#fairness

#fundamentals

Adjusting the output of a model after the model has been run. Post-processing can be used to enforce fairness constraints without modifying models themselves.

For example, one might apply post-processing to a binary classifier by setting a classification threshold such that equality of opportunity is maintained for some attribute by checking that the true positive rate is the same for all values of that attribute.

PR AUC (area under the PR curve)

Area under the interpolated precision-recall curve , obtained by plotting (recall, precision) points for different values of the classification threshold . Depending on how it's calculated, PR AUC may be equivalent to the average precision of the model.

پراکسیس

A core, high-performance ML library of Pax . Praxis is often called the "Layer library".

Praxis contains not just the definitions for the Layer class, but most of its supporting components as well, including:

• data inputs
• configuration libraries (HParam and Fiddle )
• optimizers

Praxis provides the definitions for the Model class.

دقت، درستی

A metric for classification models that answers the following question:

When the model predicted the positive class , what percentage of the predictions were correct?

این فرمول است:

$$\text{Precision} = \frac{\text{true positives}} {\text{true positives} + \text{false positives}}$$

جایی که:

• true positive means the model correctly predicted the positive class.
• false positive means the model mistakenly predicted the positive class.

For example, suppose a model made 200 positive predictions. Of these 200 positive predictions:

• 150 were true positives.
• 50 were false positives.

در این مورد:

$$\text{Precision} = \frac{\text{150}} {\text{150} + \text{50}} = 0.75$$

Contrast with accuracy and recall .

precision-recall curve

A curve of precision vs. recall at different classification thresholds .

پیش بینی

#fundamentals

A model's output. مثلا:

• The prediction of a binary classification model is either the positive class or the negative class.
• The prediction of a multi-class classification model is one class.
• The prediction of a linear regression model is a number.

prediction bias

A value indicating how far apart the average of predictions is from the average of labels in the dataset.

Not to be confused with the bias term in machine learning models or with bias in ethics and fairness .

predictive ML

Any traditional ("classic") machine learning system.

The term predictive ML doesn't have a formal definition. Rather, the term distinguishes a category of ML systems not based on generative AI .

predictive parity

#fairness

A fairness metric that checks whether, for a given classifier, the precision rates are equivalent for subgroups under consideration.

For example, a model that predicts college acceptance would satisfy predictive parity for nationality if its precision rate is the same for Lilliputians and Brobdingnagians.

Predictive parity is sometime also called predictive rate parity .

See "Fairness Definitions Explained" (section 3.2.1) for a more detailed discussion of predictive parity.

predictive rate parity

#fairness

Another name for predictive parity .

پیش پردازش

#fairness

Processing data before it's used to train a model. Preprocessing could be as simple as removing words from an English text corpus that don't occur in the English dictionary, or could be as complex as re-expressing data points in a way that eliminates as many attributes that are correlated with sensitive attributes as possible . Preprocessing can help satisfy fairness constraints .

pre-trained model

#زبان

#image

#generativeAI

Models or model components (such as an embedding vector ) that have been already been trained. Sometimes, you'll feed pre-trained embedding vectors into a neural network . Other times, your model will train the embedding vectors themselves rather than rely on the pre-trained embeddings.

The term pre-trained language model refers to a large language model that has gone through pre-training .

قبل از آموزش

#زبان

#image

#generativeAI

The initial training of a model on a large dataset. Some pre-trained models are clumsy giants and must typically be refined through additional training. For example, ML experts might pre-train a large language model on a vast text dataset, such as all the English pages in Wikipedia. Following pre-training, the resulting model might be further refined through any of the following techniques:

• تقطیر
• تنظیم دقیق
• instruction tuning
• parameter-efficient tuning
• prompt-tuning

prior belief

What you believe about the data before you begin training on it. For example, L ₂ regularization relies on a prior belief that weights should be small and normally distributed around zero.

probabilistic regression model

A regression model that uses not only the weights for each feature , but also the uncertainty of those weights. A probabilistic regression model generates a prediction and the uncertainty of that prediction. For example, a probabilistic regression model might yield a prediction of 325 with a standard deviation of 12. For more information about probabilistic regression models, see this Colab on tensorflow.org .

سریع

#زبان

#generativeAI

Any text entered as input to a large language model to condition the model to behave in a certain way. Prompts can be as short as a phrase or arbitrarily long (for example, the entire text of a novel). Prompts fall into multiple categories, including those shown in the following table:

Prompt category	مثال	یادداشت
سوال	How fast can a pigeon fly?
دستورالعمل	Write a funny poem about arbitrage.	A prompt that asks the large language model to do something.
مثال	Translate Markdown code to HTML. مثلا: Markdown: * list item HTML: <ul> <li>list item</li> </ul>	The first sentence in this example prompt is an instruction. The remainder of the prompt is the example.
نقش	Explain why gradient descent is used in machine learning training to a PhD in Physics.	The first part of the sentence is an instruction; the phrase "to a PhD in Physics" is the role portion.
Partial input for the model to complete	The Prime Minister of the United Kingdom lives at	A partial input prompt can either end abruptly (as this example does) or end with an underscore.

A generative AI model can respond to a prompt with text, code, images, embeddings , videos…almost anything.

prompt-based learning

#زبان

#generativeAI

A capability of certain models that enables them to adapt their behavior in response to arbitrary text input ( prompts ). In a typical prompt-based learning paradigm, a large language model responds to a prompt by generating text. For example, suppose a user enters the following prompt:

Summarize Newton's Third Law of Motion.

A model capable of prompt-based learning isn't specifically trained to answer the previous prompt. Rather, the model "knows" a lot of facts about physics, a lot about general language rules, and a lot about what constitutes generally useful answers. That knowledge is sufficient to provide a (hopefully) useful answer. Additional human feedback ("That answer was too complicated." or "What's a reaction?") enables some prompt-based learning systems to gradually improve the usefulness of their answers.

prompt design

#زبان

#generativeAI

Synonym for prompt engineering .

مهندسی سریع

#زبان

#generativeAI

The art of creating prompts that elicit the desired responses from a large language model . Humans perform prompt engineering. Writing well-structured prompts is an essential part of ensuring useful responses from a large language model. Prompt engineering depends on many factors, including:

• The dataset used to pre-train and possibly fine-tune the large language model.
• The temperature and other decoding parameters that the model uses to generate responses.

See Introduction to prompt design for more details on writing helpful prompts.

Prompt design is a synonym for prompt engineering.

prompt tuning

#زبان

#generativeAI

A parameter efficient tuning mechanism that learns a "prefix" that the system prepends to the actual prompt .

One variation of prompt tuning—sometimes called prefix tuning —is to prepend the prefix at every layer . In contrast, most prompt tuning only adds a prefix to the input layer .

Click the icon to learn more about prefixes.

For prompt tuning, the "prefix" (also known as a "soft prompt") is a handful of learned, task-specific vectors prepended to the text token embeddings from the actual prompt. The system learns the soft prompt by freezing all other model parameters and fine-tuning on a specific task.

---

proxy labels

#fundamentals

Data used to approximate labels not directly available in a dataset.

For example, suppose you must train a model to predict employee stress level. Your dataset contains a lot of predictive features but doesn't contain a label named stress level. Undaunted, you pick "workplace accidents" as a proxy label for stress level. After all, employees under high stress get into more accidents than calm employees. یا آنها؟ Maybe workplace accidents actually rise and fall for multiple reasons.

As a second example, suppose you want is it raining? to be a Boolean label for your dataset, but your dataset doesn't contain rain data. If photographs are available, you might establish pictures of people carrying umbrellas as a proxy label for is it raining? Is that a good proxy label? Possibly, but people in some cultures may be more likely to carry umbrellas to protect against sun than the rain.

Proxy labels are often imperfect. When possible, choose actual labels over proxy labels. That said, when an actual label is absent, pick the proxy label very carefully, choosing the least horrible proxy label candidate.

proxy (sensitive attributes)

#fairness

An attribute used as a stand-in for a sensitive attribute . For example, an individual's postal code might be used as a proxy for their income, race, or ethnicity.

pure function

A function whose outputs are based only on its inputs, and that has no side effects. Specifically, a pure function doesn't use or change any global state, such as the contents of a file or the value of a variable outside the function.

Pure functions can be used to create thread-safe code, which is beneficial when sharding model code across multiple accelerator chips .

JAX's function transformation methods require that the input functions are pure functions.

س

تابع Q

#rl

In reinforcement learning , the function that predicts the expected return from taking an action in a state and then following a given policy .

Q-function is also known as state-action value function .

یادگیری کیو

#rl

In reinforcement learning , an algorithm that allows an agent to learn the optimal Q-function of a Markov decision process by applying the Bellman equation . The Markov decision process models an environment .

چندک

Each bucket in quantile bucketing .

quantile bucketing

Distributing a feature's values into buckets so that each bucket contains the same (or almost the same) number of examples. For example, the following figure divides 44 points into 4 buckets, each of which contains 11 points. In order for each bucket in the figure to contain the same number of points, some buckets span a different width of x-values.

quantization

Overloaded term that could be used in either of two ways:

• Implementing quantile bucketing on a particular feature .
• Transforming data into zeroes and ones for quicker storing, training, and inferring. As Boolean data is more robust to noise and errors than other formats, quantization can improve model correctness. Quantization techniques include rounding, truncating, and binning .

صف

#TensorFlow

A TensorFlow Operation that implements a queue data structure. Typically used in I/O.

آر

جنگل تصادفی

#df

An ensemble of decision trees in which each decision tree is trained with a specific random noise, such as bagging .

Random forests are a type of decision forest .

سیاست تصادفی

#rl

In reinforcement learning , a policy that chooses an action at random.

رتبه بندی

A type of supervised learning whose objective is to order a list of items.

rank (ordinality)

The ordinal position of a class in a machine learning problem that categorizes classes from highest to lowest. For example, a behavior ranking system could rank a dog's rewards from highest (a steak) to lowest (wilted kale).

rank (Tensor)

#TensorFlow

The number of dimensions in a Tensor . For instance, a scalar has rank 0, a vector has rank 1, and a matrix has rank 2.

Not to be confused with rank (ordinality) .

ارزیاب

#fundamentals

A human who provides labels for examples . "Annotator" is another name for rater.

به خاطر آوردن

A metric for classification models that answers the following question:

When ground truth was the positive class , what percentage of predictions did the model correctly identify as the positive class?

این فرمول است:

\[\text{Recall} = \frac{\text{true positives}} {\text{true positives} + \text{false negatives}} \]

جایی که:

• true positive means the model correctly predicted the positive class.
• false negative means that the model mistakenly predicted the negative class .

For instance, suppose your model made 200 predictions on examples for which ground truth was the positive class. Of these 200 predictions:

• 180 were true positives.
• 20 were false negatives.

در این مورد:

\[\text{Recall} = \frac{\text{180}} {\text{180} + \text{20}} = 0.9 \]

Click the icon for notes about class-imbalanced datasets.

Recall is particularly useful for determining the predictive power of classification models in which the positive class is rare. For example, consider a class-imbalanced dataset in which the positive class for a certain disease occurs in only 10 patients out of a million. Suppose your model makes five million predictions that yield the following outcomes:

• 30 True Positives
• 20 False Negatives
• 4,999,000 True Negatives
• 950 False Positives

The recall of this model is therefore:

recall = TP / (TP + FN)
recall = 30 / (30 + 20) = 0.6 = 60%

By contrast, the accuracy of this model is:

accuracy = (TP + TN) / (TP + TN + FP + FN)
accuracy = (30 + 4,999,000) / (30 + 4,999,000 + 950 + 20) = 99.98%

That high value of accuracy looks impressive but is essentially meaningless. Recall is a much more useful metric for class-imbalanced datasets than accuracy.

---

سیستم توصیه

#recsystems

A system that selects for each user a relatively small set of desirable items from a large corpus. For example, a video recommendation system might recommend two videos from a corpus of 100,000 videos, selecting Casablanca and The Philadelphia Story for one user, and Wonder Woman and Black Panther for another. A video recommendation system might base its recommendations on factors such as:

• Movies that similar users have rated or watched.
• Genre, directors, actors, target demographic...

Rectified Linear Unit (ReLU)

#fundamentals

An activation function with the following behavior:

• If input is negative or zero, then the output is 0.
• If input is positive, then the output is equal to the input.

مثلا:

• If the input is -3, then the output is 0.
• If the input is +3, then the output is 3.0.

Here is a plot of ReLU:

ReLU is a very popular activation function. Despite its simple behavior, ReLU still enables a neural network to learn nonlinear relationships between features and the label .

شبکه عصبی مکرر

#seq

A neural network that is intentionally run multiple times, where parts of each run feed into the next run. Specifically, hidden layers from the previous run provide part of the input to the same hidden layer in the next run. Recurrent neural networks are particularly useful for evaluating sequences, so that the hidden layers can learn from previous runs of the neural network on earlier parts of the sequence.

For example, the following figure shows a recurrent neural network that runs four times. Notice that the values learned in the hidden layers from the first run become part of the input to the same hidden layers in the second run. Similarly, the values learned in the hidden layer on the second run become part of the input to the same hidden layer in the third run. In this way, the recurrent neural network gradually trains and predicts the meaning of the entire sequence rather than just the meaning of individual words.

مدل رگرسیون

#fundamentals

Informally, a model that generates a numerical prediction. (In contrast, a classification model generates a class prediction.) For example, the following are all regression models:

• A model that predicts a certain house's value, such as 423,000 Euros.
• A model that predicts a certain tree's life expectancy, such as 23.2 years.
• A model that predicts the amount of rain that will fall in a certain city over the next six hours, such as 0.18 inches.

Two common types of regression models are:

• Linear regression , which finds the line that best fits label values to features.
• Logistic regression , which generates a probability between 0.0 and 1.0 that a system typically then maps to a class prediction.

Not every model that outputs numerical predictions is a regression model. In some cases, a numeric prediction is really just a classification model that happens to have numeric class names. For example, a model that predicts a numeric postal code is a classification model, not a regression model.

منظم سازی

#fundamentals

Any mechanism that reduces overfitting . Popular types of regularization include:

• L ₁ regularization
• L ₂ regularization
• dropout regularization
• early stopping (this is not a formal regularization method, but can effectively limit overfitting)

Regularization can also be defined as the penalty on a model's complexity.

Click the icon for additional notes.

Regularization is counterintuitive. Increasing regularization usually increases training loss, which is confusing because, well, isn't the goal to minimize training loss?

قطعا نه. The goal isn't to minimize training loss. The goal is to make excellent predictions on real-world examples. Remarkably, even though increasing regularization increases training loss, it usually helps models make better predictions on real-world examples.

---

regularization rate

#fundamentals

A number that specifies the relative importance of regularization during training. Raising the regularization rate reduces overfitting but may reduce the model's predictive power. Conversely, reducing or omitting the regularization rate increases overfitting.

Click the icon to see the math.

The regularization rate is usually represented as the Greek letter lambda. The following simplified loss equation shows lambda's influence:

$$\text{minimize(loss function + }\lambda\text{(regularization))}$$

where regularization is any regularization mechanism, including;

• L ₁ regularization
• L ₂ regularization

---

reinforcement learning (RL)

#rl

A family of algorithms that learn an optimal policy , whose goal is to maximize return when interacting with an environment . For example, the ultimate reward of most games is victory. Reinforcement learning systems can become expert at playing complex games by evaluating sequences of previous game moves that ultimately led to wins and sequences that ultimately led to losses.

Reinforcement Learning from Human Feedback (RLHF)

#generativeAI

#rl

Using feedback from human raters to improve the quality of a model's responses. For example, an RLHF mechanism can ask users to rate the quality of a model's response with a 👍 or 👎 emoji. The system can then adjust its future responses based on that feedback.

ReLU

#fundamentals

Abbreviation for Rectified Linear Unit .

replay buffer

#rl

In DQN -like algorithms, the memory used by the agent to store state transitions for use in experience replay .

المثنی، کپی دقیق

A copy of the training set or model , typically on another machine. For example, a system could use the following strategy for implementing data parallelism :

1. Place replicas of an existing model on multiple machines.
2. Send different subsets of the training set to each replica.
3. Aggregate the parameter updates.

reporting bias

#fairness

The fact that the frequency with which people write about actions, outcomes, or properties is not a reflection of their real-world frequencies or the degree to which a property is characteristic of a class of individuals. Reporting bias can influence the composition of data that machine learning systems learn from.

For example, in books, the word laughed is more prevalent than breathed . A machine learning model that estimates the relative frequency of laughing and breathing from a book corpus would probably determine that laughing is more common than breathing.

نمایندگی

The process of mapping data to useful features .

re-ranking

#recsystems

The final stage of a recommendation system , during which scored items may be re-graded according to some other (typically, non-ML) algorithm. Re-ranking evaluates the list of items generated by the scoring phase, taking actions such as:

• Eliminating items that the user has already purchased.
• Boosting the score of fresher items.

نسل افزوده شده با بازیابی

#fundamentals

A software architecture commonly used in large language model (LLM) applications. Common motivations to use retrieval-augmented generation include:

• Increasing the factual accuracy of the model's generated responses
• Giving the model access to knowledge it was not trained on
• Changing what knowledge the model uses
• Enabling the model to cite sources

For example, suppose that a chemistry app uses the PaLM API to generate summaries related to user queries. When the app's backend receives a query, the backend first searches for ("retrieves") data that's relevant to the user's query, appends ("augments") the relevant chemistry data to the user's query, and instructs the LLM to create a summary based on the appended data.

برگشت

#rl

In reinforcement learning, given a certain policy and a certain state, the return is the sum of all rewards that the agent expects to receive when following the policy from the state to the end of the episode . The agent accounts for the delayed nature of expected rewards by discounting rewards according to the state transitions required to obtain the reward.

Therefore, if the discount factor is \(\gamma\), and \(r_0, \ldots, r_{N}\)denote the rewards until the end of the episode, then the return calculation is as follows:

$$\text{Return} = r_0 + \gamma r_1 + \gamma^2 r_2 + \ldots + \gamma^{N-1} r_{N-1}$$

جایزه

#rl

In reinforcement learning, the numerical result of taking an action in a state , as defined by the environment .

ridge regularization

Synonym for L ₂ regularization . The term ridge regularization is more frequently used in pure statistics contexts, whereas L ₂ regularization is used more often in machine learning.

RNN

#seq

Abbreviation for recurrent neural networks .

ROC (receiver operating characteristic) Curve

#fundamentals

A graph of true positive rate vs. false positive rate for different classification thresholds in binary classification.

The shape of an ROC curve suggests a binary classification model's ability to separate positive classes from negative classes. Suppose, for example, that a binary classification model perfectly separates all the negative classes from all the positive classes:

The ROC curve for the preceding model looks as follows:

In contrast, the following illustration graphs the raw logistic regression values for a terrible model that can't separate negative classes from positive classes at all:

The ROC curve for this model looks as follows:

Meanwhile, back in the real world, most binary classification models separate positive and negative classes to some degree, but usually not perfectly. So, a typical ROC curve falls somewhere between the two extremes:

The point on an ROC curve closest to (0.0,1.0) theoretically identifies the ideal classification threshold. However, several other real-world issues influence the selection of the ideal classification threshold. For example, perhaps false negatives cause far more pain than false positives.

A numerical metric called AUC summarizes the ROC curve into a single floating-point value.

role prompting

#زبان

#generativeAI

An optional part of a prompt that identifies a target audience for a generative AI model's response. Without a role prompt, a large language model provides an answer that may or may not be useful for the person asking the questions. With a role prompt, a large language model can answer in a way that's more appropriate and more helpful for a specific target audience. For example, the role prompt portion of the following prompts are in boldface:

• Summarize this article for a PhD in economics .
• Describe how tides work for a ten-year old .
• Explain the 2008 financial crisis. Speak as you might to a young child, or a golden retriever.

ریشه

#df

The starting node (the first condition ) in a decision tree . By convention, diagrams put the root at the top of the decision tree. مثلا:

دایرکتوری ریشه

#TensorFlow

The directory you specify for hosting subdirectories of the TensorFlow checkpoint and events files of multiple models.

ریشه میانگین مربعات خطا (RMSE)

#fundamentals

The square root of the Mean Squared Error .

rotational invariance

#image

In an image classification problem, an algorithm's ability to successfully classify images even when the orientation of the image changes. For example, the algorithm can still identify a tennis racket whether it is pointing up, sideways, or down. Note that rotational invariance is not always desirable; for example, an upside-down 9 should not be classified as a 9.

See also translational invariance and size invariance .

R-squared

A regression metric indicating how much variation in a label is due to an individual feature or to a feature set. R-squared is a value between 0 and 1, which you can interpret as follows:

• An R-squared of 0 means that none of a label's variation is due to the feature set.
• An R-squared of 1 means that all of a label's variation is due to the feature set.
• An R-squared between 0 and 1 indicates the extent to which the label's variation can be predicted from a particular feature or the feature set. For example, an R-squared of 0.10 means that 10 percent of the variance in the label is due to the feature set, an R-squared of 0.20 means that 20 percent is due to the feature set, and so on.

R-squared is the square of the Pearson correlation coefficient between the values that a model predicted and ground truth .

اس

سوگیری نمونه گیری

#fairness

See selection bias .

sampling with replacement

#df

A method of picking items from a set of candidate items in which the same item can be picked multiple times. The phrase "with replacement" means that after each selection, the selected item is returned to the pool of candidate items. The inverse method, sampling without replacement , means that a candidate item can only be picked once.

For example, consider the following fruit set:

fruit = {kiwi, apple, pear, fig, cherry, lime, mango}

Suppose that the system randomly picks fig as the first item. If using sampling with replacement, then the system picks the second item from the following set:

fruit = {kiwi, apple, pear, fig, cherry, lime, mango}

Yes, that's the same set as before, so the system could potentially pick fig again.

If using sampling without replacement, once picked, a sample can't be picked again. For example, if the system randomly picks fig as the first sample, then fig can't be picked again. Therefore, the system picks the second sample from the following (reduced) set:

fruit = {kiwi, apple, pear, cherry, lime, mango}

Click the icon for additional notes.

The word replacement in sampling with replacement confuses many people. In English, replacement means "substitution." However, sampling with replacement actually uses the French definition for replacement , which means "putting something back." The English word replacement is translated as the French word remplacement .

---

SavedModel

#TensorFlow

The recommended format for saving and recovering TensorFlow models. SavedModel is a language-neutral, recoverable serialization format, which enables higher-level systems and tools to produce, consume, and transform TensorFlow models.

See the Saving and Restoring chapter in the TensorFlow Programmer's Guide for complete details.

پس انداز

#TensorFlow

A TensorFlow object responsible for saving model checkpoints.

اسکالر

A single number or a single string that can be represented as a tensor of rank 0. For example, the following lines of code each create one scalar in TensorFlow:

breed = tf.Variable("poodle", tf.string)
temperature = tf.Variable(27, tf.int16)
precision = tf.Variable(0.982375101275, tf.float64)

پوسته پوسته شدن

Any mathematical transform or technique that shifts the range of a label and/or feature value. Some forms of scaling are very useful for transformations like normalization .

Common forms of scaling useful in Machine Learning include:

• linear scaling, which typically uses a combination of subtraction and division to replace the original value with a number between -1 and +1 or between 0 and 1.
• logarithmic scaling, which replaces the original value with its logarithm.
• Z-score normalization , which replaces the original value with a floating-point value representing the number of standard deviations from that feature's mean.

scikit-یادگیری

A popular open-source machine learning platform. See scikit-learn.org .

به ثمر رساندن

#recsystems

The part of a recommendation system that provides a value or ranking for each item produced by the candidate generation phase.

سوگیری انتخاب

#fairness

Errors in conclusions drawn from sampled data due to a selection process that generates systematic differences between samples observed in the data and those not observed. The following forms of selection bias exist:

• coverage bias : The population represented in the dataset doesn't match the population that the machine learning model is making predictions about.
• sampling bias : Data is not collected randomly from the target group.
• non-response bias (also called participation bias ): Users from certain groups opt-out of surveys at different rates than users from other groups.

For example, suppose you are creating a machine learning model that predicts people's enjoyment of a movie. To collect training data, you hand out a survey to everyone in the front row of a theater showing the movie. Offhand, this may sound like a reasonable way to gather a dataset; however, this form of data collection may introduce the following forms of selection bias:

• coverage bias: By sampling from a population who chose to see the movie, your model's predictions may not generalize to people who did not already express that level of interest in the movie.
• sampling bias: Rather than randomly sampling from the intended population (all the people at the movie), you sampled only the people in the front row. It is possible that the people sitting in the front row were more interested in the movie than those in other rows.
• non-response bias: In general, people with strong opinions tend to respond to optional surveys more frequently than people with mild opinions. Since the movie survey is optional, the responses are more likely to form a bimodal distribution than a normal (bell-shaped) distribution.

self-attention (also called self-attention layer)

#زبان

A neural network layer that transforms a sequence of embeddings (for instance, token embeddings) into another sequence of embeddings. Each embedding in the output sequence is constructed by integrating information from the elements of the input sequence through an attention mechanism.

The self part of self-attention refers to the sequence attending to itself rather than to some other context. Self-attention is one of the main building blocks for Transformers and uses dictionary lookup terminology, such as “query”, “key”, and “value”.

A self-attention layer starts with a sequence of input representations, one for each word. The input representation for a word can be a simple embedding. For each word in an input sequence, the network scores the relevance of the word to every element in the whole sequence of words. The relevance scores determine how much the word's final representation incorporates the representations of other words.

برای مثال جمله زیر را در نظر بگیرید:

The animal didn't cross the street because it was too tired.

The following illustration (from Transformer: A Novel Neural Network Architecture for Language Understanding ) shows a self-attention layer's attention pattern for the pronoun it , with the darkness of each line indicating how much each word contributes to the representation:

The self-attention layer highlights words that are relevant to "it". In this case, the attention layer has learned to highlight words that it might refer to, assigning the highest weight to animal .

For a sequence of n tokens , self-attention transforms a sequence of embeddings n separate times, once at each position in the sequence.

Refer also to attention and multi-head self-attention .

self-supervised learning

A family of techniques for converting an unsupervised machine learning problem into a supervised machine learning problem by creating surrogate labels from unlabeled examples .

Some Transformer -based models such as BERT use self-supervised learning.

Self-supervised training is a semi-supervised learning approach.

خود آموزی

A variant of self-supervised learning that is particularly useful when all of the following conditions are true:

• The ratio of unlabeled examples to labeled examples in the dataset is high.
• این یک مشکل طبقه بندی است.

Self-training works by iterating over the following two steps until the model stops improving:

1. Use supervised machine learning to train a model on the labeled examples.
2. Use the model created in Step 1 to generate predictions (labels) on the unlabeled examples, moving those in which there is high confidence into the labeled examples with the predicted label.

Notice that each iteration of Step 2 adds more labeled examples for Step 1 to train on.

semi-supervised learning

Training a model on data where some of the training examples have labels but others don't. One technique for semi-supervised learning is to infer labels for the unlabeled examples, and then to train on the inferred labels to create a new model. Semi-supervised learning can be useful if labels are expensive to obtain but unlabeled examples are plentiful.

Self-training is one technique for semi-supervised learning.

sensitive attribute

#fairness

A human attribute that may be given special consideration for legal, ethical, social, or personal reasons.

تحلیل احساسات

#زبان

Using statistical or machine learning algorithms to determine a group's overall attitude—positive or negative—toward a service, product, organization, or topic. For example, using natural language understanding , an algorithm could perform sentiment analysis on the textual feedback from a university course to determine the degree to which students generally liked or disliked the course.

sequence model

#seq

A model whose inputs have a sequential dependence. For example, predicting the next video watched from a sequence of previously watched videos.

sequence-to-sequence task

#زبان

A task that converts an input sequence of tokens to an output sequence of tokens. For example, two popular kinds of sequence-to-sequence tasks are:

• Translators:
  • Sample input sequence: "I love you."
  • Sample output sequence: "Je t'aime."
• Question answering:
  • Sample input sequence: "Do I need my car in New York City?"
  • Sample output sequence: "No. Please keep your car at home."

خدمت کردن

A synonym for inferring .

shape (Tensor)

The number of elements in each dimension of a tensor. The shape is represented as a list of integers. For example, the following two-dimensional tensor has a shape of [3,4]:

[[5, 7, 6, 4],
 [2, 9, 4, 8],
 [3, 6, 5, 1]]

TensorFlow uses row-major (C-style) format to represent the order of dimensions, which is why the shape in TensorFlow is [3,4] rather than [4,3]. In other words, in a two-dimensional TensorFlow Tensor, the shape is [ number of rows , number of columns ].

تکه شکسته

#TensorFlow

#GoogleCloud

A logical division of the training set or the model . Typically, some process creates shards by dividing the examples or parameters into (usually) equal-sized chunks. Each shard is then assigned to a different machine.

Sharding a model is called model parallelism ; sharding data is called data parallelism .

انقباض

#df

A hyperparameter in gradient boosting that controls overfitting . Shrinkage in gradient boosting is analogous to learning rate in gradient descent . Shrinkage is a decimal value between 0.0 and 1.0. A lower shrinkage value reduces overfitting more than a larger shrinkage value.

sigmoid function

#fundamentals

A mathematical function that "squishes" an input value into a constrained range, typically 0 to 1 or -1 to +1. That is, you can pass any number (two, a million, negative billion, whatever) to a sigmoid and the output will still be in the constrained range. A plot of the sigmoid activation function looks as follows:

The sigmoid function has several uses in machine learning, including:

• Converting the raw output of a logistic regression or multinomial regression model to a probability.
• Acting as an activation function in some neural networks.

Click the icon to see the math.

The sigmoid function over an input number x has the following formula:

$$ sigmoid(x) = \frac{1}{1 + e^{-\text{x}}} $$

In machine learning, x is generally a weighted sum .

---

similarity measure

#clustering

In clustering algorithms, the metric used to determine how alike (how similar) any two examples are.

single program / multiple data (SPMD)

A parallelism technique where the same computation is run on different input data in parallel on different devices. The goal of SPMD is to obtain results more quickly. It is the most common style of parallel programming.

size invariance

#image

In an image classification problem, an algorithm's ability to successfully classify images even when the size of the image changes. For example, the algorithm can still identify a cat whether it consumes 2M pixels or 200K pixels. Note that even the best image classification algorithms still have practical limits on size invariance. For example, an algorithm (or human) is unlikely to correctly classify a cat image consuming only 20 pixels.

See also translational invariance and rotational invariance .

طراحی

#clustering

In unsupervised machine learning , a category of algorithms that perform a preliminary similarity analysis on examples. Sketching algorithms use a locality-sensitive hash function to identify points that are likely to be similar, and then group them into buckets.

Sketching decreases the computation required for similarity calculations on large datasets. Instead of calculating similarity for every single pair of examples in the dataset, we calculate similarity only for each pair of points within each bucket.

سافت مکس

#fundamentals

A function that determines probabilities for each possible class in a multi-class classification model . The probabilities add up to exactly 1.0. For example, the following table shows how softmax distributes various probabilities:

Image is a...	احتمال
سگ	.85
گربه	.13
اسب	.02

Softmax is also called full softmax .

Contrast with candidate sampling .

Click the icon to see the math.

The softmax equation is as follows:

$$\sigma_i = \frac{e^{\text{z}_i}} {\sum_{j=1}^{j=K} {e^{\text{z}_j}}} $$

جایی که:

• $\sigma_i$ is the output vector. Each element of the output vector specifies the probability of this element. The sum of all the elements in the output vector is 1.0. The output vector contains the same number of elements as the input vector, $z$.
• $z$ is the input vector. Each element of the input vector contains a floating-point value.
• $K$ is the number of elements in the input vector (and the output vector).

For example, suppose the input vector is:

[1.2, 2.5, 1.8]

Therefore, softmax calculates the denominator as follows:

$$\text{denominator} = e^{1.2} + e^{2.5} + e^{1.8} = 21.552$$

The softmax probability of each element is therefore:

$$\sigma_1 = \frac{e^{1.2}}{21.552} = 0.154 $$$$\sigma_2 = \frac{e^{2.5}}{21.552} = 0.565 $$$$\sigma_1 = \frac{e^{1.8}}{21.552} = 0.281 $$

So, the output vector is therefore:

$$\sigma = [0.154, 0.565, 0.281]$$

The sum of the three elements in $\sigma$ is 1.0. اوه!

---

sparse feature

#زبان

#fundamentals

A feature whose values are predominately zero or empty. For example, a feature containing a single 1 value and a million 0 values is sparse. In contrast, a dense feature has values that are predominantly not zero or empty.

In machine learning, a surprising number of features are sparse features. Categorical features are usually sparse features. For example, of the 300 possible tree species in a forest, a single example might identify just a maple tree . Or, of the millions of possible videos in a video library, a single example might identify just "Casablanca."

In a model, you typically represent sparse features with one-hot encoding . If the one-hot encoding is big, you might put an embedding layer on top of the one-hot encoding for greater efficiency.

sparse representation

#زبان

#fundamentals

Storing only the position(s) of nonzero elements in a sparse feature.

For example, suppose a categorical feature named species identifies the 36 tree species in a particular forest. Further assume that each example identifies only a single species.

You could use a one-hot vector to represent the tree species in each example. A one-hot vector would contain a single 1 (to represent the particular tree species in that example) and 35 0 s (to represent the 35 tree species not in that example). So, the one-hot representation of maple might look something like the following:

Alternatively, sparse representation would simply identify the position of the particular species. If maple is at position 24, then the sparse representation of maple would simply be:

24

Notice that the sparse representation is much more compact than the one-hot representation.

Note: You shouldn't pass a sparse representation as a direct feature input to a model. Instead, you should convert the sparse representation into a one-hot representation before training on it.

Click the icon for a slightly more complex example.

Suppose each example in your model must represent the words—but not the order of those words—in an English sentence. English consists of about 170,000 words, so English is a categorical feature with about 170,000 elements. Most English sentences use an extremely tiny fraction of those 170,000 words, so the set of words in a single example is almost certainly going to be sparse data.

جمله زیر را در نظر بگیرید:

My dog is a great dog

You could use a variant of one-hot vector to represent the words in this sentence. In this variant, multiple cells in the vector can contain a nonzero value. Furthermore, in this variant, a cell can contain an integer other than one. Although the words "my", "is", "a", and "great" appear only once in the sentence, the word "dog" appears twice. Using this variant of one-hot vectors to represent the words in this sentence yields the following 170,000-element vector:

A sparse representation of the same sentence would simply be:

0: 1
26100: 2
45770: 1
58906: 1
91520: 1

---

Click the icon if you are confused.

The term "sparse representation" confuses a lot of people because sparse representation is itself not a sparse vector . Rather, sparse representation is actually a dense representation of a sparse vector . The synonym index representation is a little clearer than "sparse representation."

---

sparse vector

#fundamentals

A vector whose values are mostly zeroes. See also sparse feature and sparsity .

پراکندگی

The number of elements set to zero (or null) in a vector or matrix divided by the total number of entries in that vector or matrix. For example, consider a 100-element matrix in which 98 cells contain zero. The calculation of sparsity is as follows:

$$ {\text{sparsity}} = \frac{\text{98}} {\text{100}} = {\text{0.98}} $$

Feature sparsity refers to the sparsity of a feature vector; model sparsity refers to the sparsity of the model weights.

spatial pooling

#image

See pooling .

شکاف

#df

In a decision tree , another name for a condition .

شکافنده

#df

While training a decision tree , the routine (and algorithm) responsible for finding the best condition at each node .

SPMD

Abbreviation for single program / multiple data .

squared hinge loss

The square of the hinge loss . Squared hinge loss penalizes outliers more harshly than regular hinge loss.

squared loss

#fundamentals

Synonym for L ₂ loss .

staged training

#زبان

A tactic of training a model in a sequence of discrete stages. The goal can be either to speed up the training process, or to achieve better model quality.

An illustration of the progressive stacking approach is shown below:

• Stage 1 contains 3 hidden layers, stage 2 contains 6 hidden layers, and stage 3 contains 12 hidden layers.
• Stage 2 begins training with the weights learned in the 3 hidden layers of Stage 1. Stage 3 begins training with the weights learned in the 6 hidden layers of Stage 2.

See also pipelining .

حالت

#rl

In reinforcement learning, the parameter values that describe the current configuration of the environment, which the agent uses to choose an action .

state-action value function

#rl

Synonym for Q-function .

ایستا

#fundamentals

Something done once rather than continuously. The terms static and offline are synonyms. The following are common uses of static and offline in machine learning:

• static model (or offline model ) is a model trained once and then used for a while.
• static training (or offline training ) is the process of training a static model.
• static inference (or offline inference ) is a process in which a model generates a batch of predictions at a time.

Contrast with dynamic .

static inference

#fundamentals

Synonym for offline inference .

ایستایی

#fundamentals

A feature whose values don't change across one or more dimensions, usually time. For example, a feature whose values look about the same in 2021 and 2023 exhibits stationarity.

In the real world, very few features exhibit stationarity. Even features synonymous with stability (like sea level) change over time.

Contrast with nonstationarity .

گام

A forward pass and backward pass of one batch .

See backpropagation for more information on the forward pass and backward pass.

اندازه گام

Synonym for learning rate .

stochastic gradient descent (SGD)

#fundamentals

A gradient descent algorithm in which the batch size is one. In other words, SGD trains on a single example chosen uniformly at random from a training set .

گام های بلند برداشتن

#image

In a convolutional operation or pooling, the delta in each dimension of the next series of input slices. For example, the following animation demonstrates a (1,1) stride during a convolutional operation. Therefore, the next input slice starts one position to the right of the previous input slice. When the operation reaches the right edge, the next slice is all the way over to the left but one position down.

The preceding example demonstrates a two-dimensional stride. If the input matrix is three-dimensional, the stride would also be three-dimensional.

structural risk minimization (SRM)

An algorithm that balances two goals:

• The desire to build the most predictive model (for example, lowest loss).
• The desire to keep the model as simple as possible (for example, strong regularization).

For example, a function that minimizes loss+regularization on the training set is a structural risk minimization algorithm.

Contrast with empirical risk minimization .

نمونه گیری فرعی

#image

See pooling .

خلاصه

#TensorFlow

In TensorFlow, a value or set of values calculated at a particular step , usually used for tracking model metrics during training.

supervised machine learning

#fundamentals

Training a model from features and their corresponding labels . Supervised machine learning is analogous to learning a subject by studying a set of questions and their corresponding answers. After mastering the mapping between questions and answers, a student can then provide answers to new (never-before-seen) questions on the same topic.

Compare with unsupervised machine learning .

synthetic feature

#fundamentals

A feature not present among the input features, but assembled from one or more of them. Methods for creating synthetic features include the following:

• Bucketing a continuous feature into range bins.
• Creating a feature cross .
• Multiplying (or dividing) one feature value by other feature value(s) or by itself. For example, if a and b are input features, then the following are examples of synthetic features:
  • ab
  • یک ²
• Applying a transcendental function to a feature value. For example, if c is an input feature, then the following are examples of synthetic features:
  • sin(c)
  • ln(c)

Features created by normalizing or scaling alone are not considered synthetic features.

تی

T5

#زبان

A text-to-text transfer learning model introduced by Google AI in 2020 . T5 is an encoder - decoder model, based on the Transformer architecture, trained on an extremely large dataset. It is effective at a variety of natural language processing tasks, such as generating text, translating languages, and answering questions in a conversational manner.

T5 gets its name from the five T's in "Text-to-Text Transfer Transformer."

T5X

#زبان

An open-source, machine learning framework designed to build and train large-scale natural language processing (NLP) models. T5 is implemented on the T5X codebase (which is built on JAX and Flax ).

tabular Q-learning

#rl

In reinforcement learning , implementing Q-learning by using a table to store the Q-functions for every combination of state and action .

هدف

Synonym for label .

target network

#rl

In Deep Q-learning , a neural network that is a stable approximation of the main neural network, where the main neural network implements either a Q-function or a policy . Then, you can train the main network on the Q-values predicted by the target network. Therefore, you prevent the feedback loop that occurs when the main network trains on Q-values predicted by itself. By avoiding this feedback, training stability increases.

وظیفه

A problem that can be solved using machine learning techniques, such as:

• طبقه بندی
• پسرفت
• خوشه بندی
• تشخیص ناهنجاری

درجه حرارت

#زبان

#image

#generativeAI

A hyperparameter that controls the degree of randomness of a model's output. Higher temperatures result in more random output, while lower temperatures result in less random output.

Choosing the best temperature depends on the specific application and the desired properties of the model's output. For example, you would probably raise the temperature when creating an application that generates creative output. Conversely, you would probably lower the temperature when building a model that classifies images or text in order to improve the model's accuracy and consistency.

Temperature is often used with softmax .

temporal data

Data recorded at different points in time. For example, winter coat sales recorded for each day of the year would be temporal data.

تانسور

#TensorFlow

The primary data structure in TensorFlow programs. Tensors are N-dimensional (where N could be very large) data structures, most commonly scalars, vectors, or matrices. The elements of a Tensor can hold integer, floating-point, or string values.

TensorBoard

#TensorFlow

The dashboard that displays the summaries saved during the execution of one or more TensorFlow programs.

TensorFlow

#TensorFlow

A large-scale, distributed, machine learning platform. The term also refers to the base API layer in the TensorFlow stack, which supports general computation on dataflow graphs.

Although TensorFlow is primarily used for machine learning, you may also use TensorFlow for non-ML tasks that require numerical computation using dataflow graphs.

TensorFlow Playground

#TensorFlow

A program that visualizes how different hyperparameters influence model (primarily neural network) training. Go to http://playground.tensorflow.org to experiment with TensorFlow Playground.

سرویس TensorFlow

#TensorFlow

A platform to deploy trained models in production.

واحد پردازش تانسور (TPU)

#TensorFlow

#GoogleCloud

An application-specific integrated circuit (ASIC) that optimizes the performance of machine learning workloads. These ASICs are deployed as multiple TPU chips on a TPU device .

Tensor rank

#TensorFlow

See rank (Tensor) .

Tensor shape

#TensorFlow

The number of elements a Tensor contains in various dimensions. For example, a [5, 10] Tensor has a shape of 5 in one dimension and 10 in another.

Tensor size

#TensorFlow

The total number of scalars a Tensor contains. For example, a [5, 10] Tensor has a size of 50.

TensorStore

A library for efficiently reading and writing large multi-dimensional arrays.

termination condition

#rl

In reinforcement learning , the conditions that determine when an episode ends, such as when the agent reaches a certain state or exceeds a threshold number of state transitions. For example, in tic-tac-toe (also known as noughts and crosses), an episode terminates either when a player marks three consecutive spaces or when all spaces are marked.

تست

#df

In a decision tree , another name for a condition .

test loss

#fundamentals

A metric representing a model's loss against the test set . When building a model , you typically try to minimize test loss. That's because a low test loss is a stronger quality signal than a low training loss or low validation loss .

A large gap between test loss and training loss or validation loss sometimes suggests that you need to increase the regularization rate .

مجموعه تست

A subset of the dataset reserved for testing a trained model .

Traditionally, you divide examples in the dataset into the following three distinct subsets:

• a training set
• a validation set
• a test set

Each example in a dataset should belong to only one of the preceding subsets. For instance, a single example should not belong to both the training set and the test set.

The training set and validation set are both closely tied to training a model. Because the test set is only indirectly associated with training, test loss is a less biased, higher quality metric than training loss or validation loss .

text span

#زبان

The array index span associated with a specific subsection of a text string. For example, the word good in the Python string s="Be good now" occupies the text span from 3 to 6.

tf.Example

#TensorFlow

A standard protocol buffer for describing input data for machine learning model training or inference.

tf.keras

#TensorFlow

An implementation of Keras integrated into TensorFlow .

threshold (for decision trees)

#df

In an axis-aligned condition , the value that a feature is being compared against. For example, 75 is the threshold value in the following condition:

grade >= 75

This form of the term threshold is different than classification threshold .

تجزیه و تحلیل سری های زمانی

#clustering

A subfield of machine learning and statistics that analyzes temporal data . Many types of machine learning problems require time series analysis, including classification, clustering, forecasting, and anomaly detection. For example, you could use time series analysis to forecast the future sales of winter coats by month based on historical sales data.

مرحله زمانی

#seq

One "unrolled" cell within a recurrent neural network . For example, the following figure shows three timesteps (labeled with the subscripts t-1, t, and t+1):

token

#زبان

In a language model , the atomic unit that the model is training on and making predictions on. A token is typically one of the following:

• a word—for example, the phrase "dogs like cats" consists of three word tokens: "dogs", "like", and "cats".
• a character—for example, the phrase "bike fish" consists of nine character tokens. (Note that the blank space counts as one of the tokens.)
• subwords—in which a single word can be a single token or multiple tokens. A subword consists of a root word, a prefix, or a suffix. For example, a language model that uses subwords as tokens might view the word "dogs" as two tokens (the root word "dog" and the plural suffix "s"). That same language model might view the single word "taller" as two subwords (the root word "tall" and the suffix "er").

In domains outside of language models, tokens can represent other kinds of atomic units. For example, in computer vision, a token might be a subset of an image.

برج

A component of a deep neural network that is itself a deep neural network without an output layer. Typically, each tower reads from an independent data source. Towers are independent until their output is combined in a final layer.

TPU

#TensorFlow

#GoogleCloud

Abbreviation for Tensor Processing Unit .

TPU chip

#TensorFlow

#GoogleCloud

A programmable linear algebra accelerator with on-chip high bandwidth memory that is optimized for machine learning workloads. Multiple TPU chips are deployed on a TPU device .

TPU device

#TensorFlow

#GoogleCloud

A printed circuit board (PCB) with multiple TPU chips , high bandwidth network interfaces, and system cooling hardware.

TPU master

#TensorFlow

#GoogleCloud

The central coordination process running on a host machine that sends and receives data, results, programs, performance, and system health information to the TPU workers . The TPU master also manages the setup and shutdown of TPU devices .

TPU node

#TensorFlow

#GoogleCloud

A TPU resource on Google Cloud Platform with a specific TPU type . The TPU node connects to your VPC Network from a peer VPC network . TPU nodes are a resource defined in the Cloud TPU API .

TPU Pod

#TensorFlow

#GoogleCloud

A specific configuration of TPU devices in a Google data center. All of the devices in a TPU Pod are connected to one another over a dedicated high-speed network. A TPU Pod is the largest configuration of TPU devices available for a specific TPU version.

TPU resource

#TensorFlow

#GoogleCloud

A TPU entity on Google Cloud Platform that you create, manage, or consume. For example, TPU nodes and TPU types are TPU resources.

TPU slice

#TensorFlow

#GoogleCloud

A TPU slice is a fractional portion of the TPU devices in a TPU Pod . All of the devices in a TPU slice are connected to one another over a dedicated high-speed network.

TPU type

#TensorFlow

#GoogleCloud

A configuration of one or more TPU devices with a specific TPU hardware version. You select a TPU type when you create a TPU node on Google Cloud Platform. For example, a v2-8 TPU type is a single TPU v2 device with 8 cores. A v3-2048 TPU type has 256 networked TPU v3 devices and a total of 2048 cores. TPU types are a resource defined in the Cloud TPU API .

TPU worker

#TensorFlow

#GoogleCloud

A process that runs on a host machine and executes machine learning programs on TPU devices .

آموزش

#fundamentals

The process of determining the ideal parameters (weights and biases) comprising a model . During training, a system reads in examples and gradually adjusts parameters. Training uses each example anywhere from a few times to billions of times.

training loss

#fundamentals

A metric representing a model's loss during a particular training iteration. For example, suppose the loss function is Mean Squared Error . Perhaps the training loss (the Mean Squared Error) for the 10th iteration is 2.2, and the training loss for the 100th iteration is 1.9.

A loss curve plots training loss vs. the number of iterations. A loss curve provides the following hints about training:

• A downward slope implies that the model is improving.
• An upward slope implies that the model is getting worse.
• A flat slope implies that the model has reached convergence .

For example, the following somewhat idealized loss curve shows:

• A steep downward slope during the initial iterations, which implies rapid model improvement.
• A gradually flattening (but still downward) slope until close to the end of training, which implies continued model improvement at a somewhat slower pace then during the initial iterations.
• A flat slope towards the end of training, which suggests convergence.

Although training loss is important, see also generalization .

training-serving skew

#fundamentals

The difference between a model's performance during training and that same model's performance during serving .

training set

#fundamentals

The subset of the dataset used to train a model .

Traditionally, examples in the dataset are divided into the following three distinct subsets:

• a training set
• a validation set
• a test set

Ideally, each example in the dataset should belong to only one of the preceding subsets. For example, a single example should not belong to both the training set and the validation set.

خط سیر

#rl

In reinforcement learning , a sequence of tuples that represent a sequence of state transitions of the agent , where each tuple corresponds to the state, action , reward , and next state for a given state transition.

انتقال یادگیری

Transferring information from one machine learning task to another. For example, in multi-task learning, a single model solves multiple tasks, such as a deep model that has different output nodes for different tasks. Transfer learning might involve transferring knowledge from the solution of a simpler task to a more complex one, or involve transferring knowledge from a task where there is more data to one where there is less data.

Most machine learning systems solve a single task. Transfer learning is a baby step towards artificial intelligence in which a single program can solve multiple tasks.

تبدیل کننده

#زبان

A neural network architecture developed at Google that relies on self-attention mechanisms to transform a sequence of input embeddings into a sequence of output embeddings without relying on convolutions or recurrent neural networks . A Transformer can be viewed as a stack of self-attention layers.

A Transformer can include any of the following:

• an encoder
• a decoder
• both an encoder and decoder

An encoder transforms a sequence of embeddings into a new sequence of the same length. An encoder includes N identical layers, each of which contains two sub-layers. These two sub-layers are applied at each position of the input embedding sequence, transforming each element of the sequence into a new embedding. The first encoder sub-layer aggregates information from across the input sequence. The second encoder sub-layer transforms the aggregated information into an output embedding.

A decoder transforms a sequence of input embeddings into a sequence of output embeddings, possibly with a different length. A decoder also includes N identical layers with three sub-layers, two of which are similar to the encoder sub-layers. The third decoder sub-layer takes the output of the encoder and applies the self-attention mechanism to gather information from it.

The blog post Transformer: A Novel Neural Network Architecture for Language Understanding provides a good introduction to Transformers.

translational invariance

#image

In an image classification problem, an algorithm's ability to successfully classify images even when the position of objects within the image changes. For example, the algorithm can still identify a dog, whether it is in the center of the frame or at the left end of the frame.

See also size invariance and rotational invariance .

سه گرام

#seq

#زبان

An N-gram in which N=3.

true negative (TN)

#fundamentals

An example in which the model correctly predicts the negative class . For example, the model infers that a particular email message is not spam , and that email message really is not spam .

true positive (TP)

#fundamentals

An example in which the model correctly predicts the positive class . For example, the model infers that a particular email message is spam, and that email message really is spam.

true positive rate (TPR)

#fundamentals

Synonym for recall . به این معنا که:

$$\text{true positive rate} = \frac{\text{true positives}} {\text{true positives} + \text{false negatives}}$$

True positive rate is the y-axis in an ROC curve .

U

unawareness (to a sensitive attribute)

#fairness

A situation in which sensitive attributes are present, but not included in the training data. Because sensitive attributes are often correlated with other attributes of one's data, a model trained with unawareness about a sensitive attribute could still have disparate impact with respect to that attribute, or violate other fairness constraints .

underfitting

#fundamentals

Producing a model with poor predictive ability because the model hasn't fully captured the complexity of the training data. Many problems can cause underfitting, including:

• Training on the wrong set of features .
• Training for too few epochs or at too low a learning rate .
• Training with too high a regularization rate .
• Providing too few hidden layers in a deep neural network.

کم نمونه گیری

Removing examples from the majority class in a class-imbalanced dataset in order to create a more balanced training set .

For example, consider a dataset in which the ratio of the majority class to the minority class is 20:1. To overcome this class imbalance, you could create a training set consisting of all of the minority class examples but only a tenth of the majority class examples, which would create a training-set class ratio of 2:1. Thanks to undersampling, this more balanced training set might produce a better model. Alternatively, this more balanced training set might contain insufficient examples to train an effective model.

Contrast with oversampling .

یک طرفه

#زبان

A system that only evaluates the text that precedes a target section of text. In contrast, a bidirectional system evaluates both the text that precedes and follows a target section of text. See bidirectional for more details.

unidirectional language model

#زبان

A language model that bases its probabilities only on the tokens appearing before , not after , the target token(s). Contrast with bidirectional language model .

unlabeled example

#fundamentals

An example that contains features but no label . For example, the following table shows three unlabeled examples from a house valuation model, each with three features but no house value:

تعداد اتاق خواب	تعداد حمام	House age
3	2	15
2	1	72
4	2	34

In supervised machine learning , models train on labeled examples and make predictions on unlabeled examples .

In semi-supervised and unsupervised learning, unlabeled examples are used during training.

Contrast unlabeled example with labeled example .

unsupervised machine learning

#clustering

#fundamentals

Training a model to find patterns in a dataset, typically an unlabeled dataset.

The most common use of unsupervised machine learning is to cluster data into groups of similar examples. For example, an unsupervised machine learning algorithm can cluster songs based on various properties of the music. The resulting clusters can become an input to other machine learning algorithms (for example, to a music recommendation service). Clustering can help when useful labels are scarce or absent. For example, in domains such as anti-abuse and fraud, clusters can help humans better understand the data.

Contrast with supervised machine learning .

Click the icon for additional notes.

Another example of unsupervised machine learning is principal component analysis (PCA) . For example, applying PCA on a dataset containing the contents of millions of shopping carts might reveal that shopping carts containing lemons frequently also contain antacids.

---

uplift modeling

A modeling technique, commonly used in marketing, that models the "causal effect" (also known as the "incremental impact") of a "treatment" on an "individual." در اینجا دو نمونه وجود دارد:

• Doctors might use uplift modeling to predict the mortality decrease (causal effect) of a medical procedure (treatment) depending on the age and medical history of a patient (individual).
• Marketers might use uplift modeling to predict the increase in probability of a purchase (causal effect) due to an advertisement (treatment) on a person (individual).

Uplift modeling differs from classification or regression in that some labels (for example, half of the labels in binary treatments) are always missing in uplift modeling. For example, a patient can either receive or not receive a treatment; therefore, we can only observe whether the patient is going to heal or not heal in only one of these two situations (but never both). The main advantage of an uplift model is that it can generate predictions for the unobserved situation (the counterfactual) and use it to compute the causal effect.

upweighting

Applying a weight to the downsampled class equal to the factor by which you downsampled.

user matrix

#recsystems

In recommendation systems , an embedding vector generated by matrix factorization that holds latent signals about user preferences. Each row of the user matrix holds information about the relative strength of various latent signals for a single user. For example, consider a movie recommendation system. In this system, the latent signals in the user matrix might represent each user's interest in particular genres, or might be harder-to-interpret signals that involve complex interactions across multiple factors.

The user matrix has a column for each latent feature and a row for each user. That is, the user matrix has the same number of rows as the target matrix that is being factorized. For example, given a movie recommendation system for 1,000,000 users, the user matrix will have 1,000,000 rows.

V

اعتبار سنجی

#fundamentals

The initial evaluation of a model's quality. Validation checks the quality of a model's predictions against the validation set .

Because the validation set differs from the training set , validation helps guard against overfitting .

You might think of evaluating the model against the validation set as the first round of testing and evaluating the model against the test set as the second round of testing.

validation loss

#fundamentals

A metric representing a model's loss on the validation set during a particular iteration of training.

See also generalization curve .

validation set

#fundamentals

The subset of the dataset that performs initial evaluation against a trained model . Typically, you evaluate the trained model against the validation set several times before evaluating the model against the test set .

Traditionally, you divide the examples in the dataset into the following three distinct subsets:

• a training set
• a validation set
• a test set

Ideally, each example in the dataset should belong to only one of the preceding subsets. For example, a single example should not belong to both the training set and the validation set.

value imputation

The process of replacing a missing value with an acceptable substitute. When a value is missing, you can either discard the entire example or you can use value imputation to salvage the example.

For example, consider a dataset containing a temperature feature that is supposed to be recorded every hour. However, the temperature reading was unavailable for a particular hour. Here is a section of the dataset:

مهر زمان	درجه حرارت
1680561000	10
1680564600	12
1680568200	گم شده
1680571800	20
1680575400	21
1680579000	21

A system could either delete the missing example or impute the missing temperature as 12, 16, 18, or 20, depending on the imputation algorithm.

vanishing gradient problem

#seq

The tendency for the gradients of early hidden layers of some deep neural networks to become surprisingly flat (low). Increasingly lower gradients result in increasingly smaller changes to the weights on nodes in a deep neural network, leading to little or no learning. Models suffering from the vanishing gradient problem become difficult or impossible to train. Long Short-Term Memory cells address this issue.

Compare to exploding gradient problem .

variable importances

#df

A set of scores that indicates the relative importance of each feature to the model.

For example, consider a decision tree that estimates house prices. Suppose this decision tree uses three features: size, age, and style. If a set of variable importances for the three features are calculated to be {size=5.8, age=2.5, style=4.7}, then size is more important to the decision tree than age or style.

Different variable importance metrics exist, which can inform ML experts about different aspects of models.

variational autoencoder (VAE)

#زبان

A type of autoencoder that leverages the discrepancy between inputs and outputs to generate modified versions of the inputs. Variational autoencoders are useful for generative AI .

VAEs are based on variational inference: a technique for estimating the parameters of a probability model.

دبلیو

Wasserstein loss

One of the loss functions commonly used in generative adversarial networks , based on the earth mover's distance between the distribution of generated data and real data.

وزن

#fundamentals

A value that a model multiplies by another value. Training is the process of determining a model's ideal weights; inference is the process of using those learned weights to make predictions.

Click the icon to see an example of weights in a linear model.

Imagine a linear model with two features. Suppose that training determines the following weights (and bias ):

• The bias, b, has a value of 2.2
• The weight, w ₁ associated with one feature is 1.5.
• The weight, w ₂ associated with the other feature is 0.4.

Now imagine an example with the following feature values:

• The value of one feature, x ₁ , is 6.
• The value of the other feature, x ₂ , is 10.

This linear model uses the following formula to generate a prediction, y':

$$y' = b + w_1x_1 + w_2x_2$$

Therefore, the prediction is:

$$y' = 2.2 + (1.5)(6) + (0.4)(10) = 15.2$$

If a weight is 0, then the corresponding feature doesn't contribute to the model. For example, if w ₁ is 0, then the value of x ₁ is irrelevant.

---

Weighted Alternating Least Squares (WALS)

#recsystems

An algorithm for minimizing the objective function during matrix factorization in recommendation systems , which allows a downweighting of the missing examples. WALS minimizes the weighted squared error between the original matrix and the reconstruction by alternating between fixing the row factorization and column factorization. Each of these optimizations can be solved by least squares convex optimization . For details, see the Recommendation Systems course .

جمع وزنی

#fundamentals

The sum of all the relevant input values multiplied by their corresponding weights. For example, suppose the relevant inputs consist of the following:

مقدار ورودی	input weight
2	-1.3
-1	0.6
3	0.4

The weighted sum is therefore:

weighted sum = (2)(-1.3) + (-1)(0.6) + (3)(0.4) = -2.0

A weighted sum is the input argument to an activation function .

wide model

A linear model that typically has many sparse input features . We refer to it as "wide" since such a model is a special type of neural network with a large number of inputs that connect directly to the output node. Wide models are often easier to debug and inspect than deep models . Although wide models cannot express nonlinearities through hidden layers , wide models can use transformations such as feature crossing and bucketization to model nonlinearities in different ways.

Contrast with deep model .

عرض

The number of neurons in a particular layer of a neural network .

حکمت جمعیت

#df

The idea that averaging the opinions or estimates of a large group of people ("the crowd") often produces surprisingly good results. For example, consider a game in which people guess the number of jelly beans packed into a large jar. Although most individual guesses will be inaccurate, the average of all the guesses has been empirically shown to be surprisingly close to the actual number of jelly beans in the jar.

Ensembles are a software analog of wisdom of the crowd. Even if individual models make wildly inaccurate predictions, averaging the predictions of many models often generates surprisingly good predictions. For example, although an individual decision tree might make poor predictions, a decision forest often makes very good predictions.

جاسازی کلمه

#زبان

Representing each word in a word set within an embedding vector ; that is, representing each word as a vector of floating-point values between 0.0 and 1.0. Words with similar meanings have more-similar representations than words with different meanings. For example, carrots , celery , and cucumbers would all have relatively similar representations, which would be very different from the representations of airplane , sunglasses , and toothpaste .

ایکس

XLA (Accelerated Linear Algebra)

An open-source machine learning compiler for GPUs, CPUs, and ML accelerators.

The XLA compiler takes models from popular ML frameworks such as PyTorch , TensorFlow , and JAX , and optimizes them for high-performance execution across different hardware platforms including GPUs, CPUs, and ML accelerators .

ز

zero-shot learning

A type of machine learning training where the model infers a prediction for a task that it was not specifically already trained on. In other words, the model is given zero task-specific training examples but asked to do inference for that task.

zero-shot prompting

#زبان

#generativeAI

A prompt that does not provide an example of how you want the large language model to respond. مثلا:

Parts of one prompt	یادداشت
What is the official currency of the specified country?	The question you want the LLM to answer.
هند:	The actual query.

The large language model might respond with any of the following:

• روپیه
• INR
• ₹
• روپیه هند
• روپیه
• The Indian rupee

All of the answers are correct, though you might prefer a particular format.

Compare and contrast zero-shot prompting with the following terms:

• one-shot prompting
• few-shot prompting

Z-score normalization

#fundamentals

A scaling technique that replaces a raw feature value with a floating-point value representing the number of standard deviations from that feature's mean. For example, consider a feature whose mean is 800 and whose standard deviation is 100. The following table shows how Z-score normalization would map the raw value to its Z-score:

Raw value	امتیاز Z
800	0
950	+1.5
575	-2.25

The machine learning model then trains on the Z-scores for that feature instead of on the raw values.