sourced story
12 June 2017Primary source · 2 sourcesWell documented

'Attention Is All You Need' introduces the Transformer

Google researchers drop recurrence entirely and let every word attend to every other word at once

On the timeline · around 12 June 2017 · Transformers and the Generative AI WaveThe Statistical and Machine Learning TurnTransformers and the Generative AI Wave'Attention Is All You Need' introduces the Transformer20112013201520182019

Quick facts

Lead author
Ashish Vaswani
Institution
Google Brain / Google Research
Key innovation
Self-attention, no recurrence or convolution

What happened

Ashish Vaswani and seven co-authors at Google Brain and Google Research published 'Attention Is All You Need,' introducing the Transformer architecture. Earlier sequence models processed text word by word in order, using recurrent connections that made long-range relationships hard to learn and training difficult to parallelize. The Transformer replaced recurrence entirely with a self-attention mechanism, which lets the model weigh every word in a sentence against every other word simultaneously to build a representation informed by the whole context at once, rather than only what came before. Because self-attention can be computed for all words in parallel rather than one step at a time, Transformers trained dramatically faster on modern GPU hardware. The model achieved a new state-of-the-art BLEU score of 41.8 on an English-to-French translation benchmark after training for just 3.5 days on eight GPUs, far less compute than prior best results required.

Why it matters

The Transformer's parallelizable attention mechanism became the shared architectural foundation for nearly every major language model that followed, including BERT, GPT, and the large language models behind ChatGPT, because it scaled to much larger datasets and models than recurrent architectures could practically train.

How we know

The original arXiv paper documents the architecture and translation benchmark results directly; Google's own research blog post on the paper independently explains the same self-attention mechanism using a worked example of resolving word ambiguity.

Sources

See something wrong? . Corrections with a source get fixed fastest.

Part of a timelineArtificial Intelligence30 events · From a wartime theory of neurons to machines that write, paint, and fold proteinsView all →