3 ms·
The "attention" mechanism (a bit of a misnomer really) is what makes transformers more complex than many other neural nets - data isn't simply flowing through t
by HarHarVeryFunny 3y ago
The "attention" mechanism (a bit of a misnomer really) is what makes transformers more complex than many other neural nets - data isn't simply flowing through the model from layer to layer, but rather it is being copied and moved around by the attention heads. The "next word" it is generating doesn't even have to be a word it has ever seen before - it may be copying it from the prompt.
- potatoman22 3y agoInteresting. Any suggestions/references for learning about attention from this perspective?
- HarHarVeryFunny 3y agoThe paper I read was this one from Catherine Olsson et al at Anthropic. https://transformer-circuits.pub/2022/in-context-learning-and-induction-heads/index.html https://transformer-circuits.pub/2022/in-context-learning-an... There's a useful article here that expands on the types of head composition and provides some illustrations. https://www.lesswrong.com/posts/TvrfY4c9eaGLeyDkE/induction-heads-illustrated https://www.lesswrong.com/posts/TvrfY4c9eaGLeyDkE/induction-...