Библиотека · CC BY-SA 4.0

Погружение вглубокое обучение

Dive into Deep Learning

Иллюстрация к книге

Выберите версию

Один и тот же текст на каждом уровне, на всех языках и во всех оформлениях, что здесь есть. Выберите, потом слушайте или скачивайте.

Слушать и смотреть

Тот же плеер, который продукт даёт для вашей книги: голос, рисунки по мере того, как они рисуются, текст.

Открытый учебник, по которому учатся в сотнях университетских курсов о нейросетях. Первая глава объясняет, что значит для программы учиться на данных: модели, функции потерь, оптимизация, виды задач обучения и как глубокое обучение пришло к успеху.

Это пробная часть: 1 из 23 (главы).

Из лекций

A stack of recordings tagged yes and no feeds a learning algorithm, which turns the knobs on a box called the model.
Knobs that learn a word

Labels set the knobs

Here is the trick the book offers. You cannot write down how you recognise the word "Alexa", but you recognise it at once. So you can do something else: collect thousands of short recordings and mark each one yourself, yes if it contains the word and no if it does not.

A vertical loss scale with an arrow pointing down labelled better, and the formula loss = − score.
The right data and a fair score

Lower is better

The third ingredient answers a question that sounds philosophical. We said that learning means getting better at a task. But who decides what "better" is? If one person proposes a change to the model, another may well argue that it made things worse.

The same price line with a third point at 3 hours, 420 dollars, above the line's 350; a dashed gap of 70 between them and the squared miss 70² = 4,900.
How much? Regression

A bill off the line

The plumber was an easy case, because two bills fixed the two knobs exactly. Real data is rarely so kind. A third bill might have been 420 dollars for 3 hours, because the pipes were older or the job was on a Sunday, things that hours alone do not capture.

The sentence Tom has dinner in Washington with Sally with a mark under each word, Ent under the names; below, a long sound wave of 48,000 samples mapped to six words.
When order matters: sequences

Aligned tags, unaligned audio

The book picks out a few special cases. The first is tagging and parsing, where every word of a text gets a mark. Here the input and output are aligned: there is one mark per word, in the same order as the words. In part-of-speech tagging, each word gets its grammatical role, such as noun or verb.

Two coffee machines: machine A showing 6, machine B showing a question mark; arrows labelled exploit to A and explore to B.
Learning by acting: reinforcement learning

Exploit or explore

A second difficulty is partial observability: what the agent sees now may not tell it where it is. Imagine your cleaning robot stuck in one of several identical cupboards in your house. From inside, they all look the same; to find which one it is in, it must remember what it saw before it went in.

A column of four games from bottom to top, backgammon, chess, Go, poker, with an upward arrow labelled harder; poker marked hidden cards.
Hidden successes, visible ones

A ladder of games

Games have long served as a yardstick of human ability, and the book traces a ladder of them. It begins with TD-Gammon, a backgammon program that learned by a form of reinforcement learning.

Как это сделано

Примеры для: a curious adult with school maths who uses ChatGPT and phone apps; add one modern instance per idea, such as a chatbot, a spam filter, a recommendation feed or a voice assistant

Источник

Лекции пересказывают с английского оригинала книгу d2l.ai (Cambridge University Press, 2023), GitHub d2l-ai/d2l-en, автор — Астон Чжан, Закари Липтон, Му Ли, Александр Смола; книга вышла под открытой лицензией, которая разрешает пересказ и перевод. Пересказ, рисунки и голос сделаны ИИ и сверены с книгой второй моделью. Как используется ИИ.

Лекции — переработка оригинала и распространяются под той же лицензией: их можно свободно распространять и перерабатывать с указанием автора и ссылкой на эту страницу: CC BY-SA 4.0

Ваша трудная книга — так же

Загрузите учебник или любую сложную книгу. Первые пять минут бесплатно.

Загрузить книгу