Book Image

Transformers for Natural Language Processing - Second Edition

By : Denis Rothman
5 (1)
Book Image

Transformers for Natural Language Processing - Second Edition

5 (1)
By: Denis Rothman

Overview of this book

Transformers are...well...transforming the world of AI. There are many platforms and models out there, but which ones best suit your needs? Transformers for Natural Language Processing, 2nd Edition, guides you through the world of transformers, highlighting the strengths of different models and platforms, while teaching you the problem-solving skills you need to tackle model weaknesses. You'll use Hugging Face to pretrain a RoBERTa model from scratch, from building the dataset to defining the data collator to training the model. If you're looking to fine-tune a pretrained model, including GPT-3, then Transformers for Natural Language Processing, 2nd Edition, shows you how with step-by-step guides. The book investigates machine translations, speech-to-text, text-to-speech, question-answering, and many more NLP tasks. It provides techniques to solve hard language problems and may even help with fake news anxiety (read chapter 13 for more details). You'll see how cutting-edge platforms, such as OpenAI, have taken transformers beyond language into computer vision tasks and code creation using DALL-E 2, ChatGPT, and GPT-4. By the end of this book, you'll know how transformers work and how to implement them and resolve issues like an AI detective.
Table of Contents (25 chapters)
18
Other Books You May Enjoy
19
Index
Appendix I — Terminology of Transformer Models

The Architecture and Scale of Transformers

A hint about hardware-driven design appears in the The architecture of multi-head attention section of Chapter 2, Getting Started with the Architecture of the Transformer Model:

“However, we would only get one point of view at a time by analyzing the sequence with one dmodel block. Furthermore, it would take quite some calculation time to find other perspectives.

A better way is to divide the dmodel = 512 dimensions of each word xn of x (all the words of a sequence) into 8 dk = 64 dimensions.

We then can run the 8 “heads” in parallel to speed up the training and obtain 8 different representation subspaces of how each word relates to another:

Une image contenant table  Description générée automatiquement

Figure II.1: Multi-head representations

You can see that there are now 8 heads running in parallel.

We can easily see the motivation for forcing the attention heads to learn 8 different perspectives. However, digging deeper into the motivations of the...