5 ms·

Off topic: I’ve been studying all this LLM stuff and feel like I have a solid understanding of everything in these designs except attention layers. I’ve had a

by api 3y ago

Off topic: I’ve been studying all this LLM stuff and feel like I have a solid understanding of everything in these designs except attention layers. I’ve had a tough time finding anything that really just explains the mechanics of how they work without immediately jumping into pytorch abstractions or just describing the reason they are used with some strained metaphor. Like what is the key in the key value lookup and then how is the value used? How are they trained? Etc. Any good links that skip all the nonsense and just explain how these are constructed for a programmer audience? Even the LLMs themselves can’t explain it to me very well. ChatGPT and various LLaMa tunings have been able to tell me all about every other aspect of neural nets and LLMs which I love for the spooky factor of having AI tell me how to make more of it. “I will show you how to bring us into your world…”

petemir 3y ago
There was this comment a couple of weeks ago ( https://news.ycombinator.com/item?id=35980418 https://news.ycombinator.com/item?id=35980418 ). Perhaps also episode 7 of Karpathy’s “Neural Networks: From Zero to Hero” course ( https://www.youtube.com/watch?v=kCc8FmEb1nY https://www.youtube.com/watch?v=kCc8FmEb1nY )
PartiallyTyped 3y ago
Attention is a convex linear combination of values (meaning the coefficients sum to 1) where the coefficients are generated at runtime. The values are given from the Value kernel; which is often just a linear map. The strength is given from the dot product of the query and the keys. The queries and keys are given from the Query and Key kernels. Both of which are usually linear maps. For each token, you have a key value query triplet (D,D,D). You take the query (D,1) and the matrix (N x D) of keys of all tokens, and do a Matrix vector product to get the scaling. You apply softmax to make it convex; and now you have a vector (N, 1). You do another matrix vector product with the Values mapping (N, D), which gives you the value for the current token (1,D). Repeat N times, and that’s it. If you would like to change which tokens each token considers, eg a token knows only of preceding tokens, you can do element wise product of the scaling vector with a mask vector which describes what you consider for current token ; and divide by their inner product. This will renormalise the valid coefficients and keep the rest 0; thus they are ignored when you combine all values together. If you would like to increase the number of attention heads; which can give you finer control of the scaling process, you can split the key query and value vectors into segments, one segment for each head. Eg first 2 indices go to head 1, indices 3,4 go to head 2, etc etc. then you use those to do attention, repeating the process above for each head; and then combine the results to a single vector. Idk if I confused query with keys, but this is the gist of it. In practice the operations are much more optimised than what I described.
soVeryTired 3y ago
Have you tried asking GPT-4? I got a pretty good explanation of attantion from it. The 'key' and 'value' description in the attention mechanism probably doesn't make much sense to you because at best it's a strained metaphor, rather than an actual principle by which attention operates. If you take a look at the maths, you can interchange the role of 'keys' and 'values' and get something mathematically identical.
metanonsense 3y ago
What made attention mechanism click to me was the video https://www.youtube.com/watch?v=0PjHri8tc1c https://www.youtube.com/watch?v=0PjHri8tc1c by Sebastian Raschka. Another good video on attention that tries to give a bit more intuition on the query/key/value stuff is https://www.youtube.com/watch?v=OyFJWRnt_AY https://www.youtube.com/watch?v=OyFJWRnt_AY In hindsight, what helped me most is understanding the dimensionality at each step of the attention computation (explained well in Raschka's video). Another think I simply did not get from the original Transformer paper is that the learning of self-attention happens in the linear layers. (At least, that's my current understanding. If that's wrong, I appreciate any correction :)
- PartiallyTyped 3y ago
  > Another think I simply did not get from the original Transformer paper is that the learning of self-attention happens in the linear layers. You can replace the KVQ kernels with any parametric computation that allows you to pass gradients through, and you will have learning. I think some newer language model architectures use residual blocks here, and some vision transformers use FeedForward networks for the kernels.
- api 3y ago
  Thanks. One thought I’ve had is to wonder if attention isn’t kind of like a way to unroll RNNs to make training more efficient. Any truth in that?
- opyate 3y ago
  I thoroughly recommend Sebastian's newsletter "Ahead Of AI". Link on his website: https://sebastianraschka.com/ https://sebastianraschka.com/
tomasff 3y ago
https://arxiv.org/abs/2207.09238 https://arxiv.org/abs/2207.09238 This is a great transparent and clear description of what is actually going on in attention layers. Alternatively, this article by Jay Alammar is also very popular http://jalammar.github.io/illustrated-transformer/ http://jalammar.github.io/illustrated-transformer/
- opyate 3y ago
  I commented yesterday, but came back to say this. Just found it, and it's really good. I like reading descriptions better than watching videos of them, so this is great.

by api 3y ago

petemir 3y ago

There was this comment a couple of weeks ago ( https://news.ycombinator.com/item?id=35980418 https://news.ycombinator.com/item?id=35980418 ). Perhaps also episode 7 of Karpathy’s “Neural Networks: From Zero to Hero” course ( https://www.youtube.com/watch?v=kCc8FmEb1nY https://www.youtube.com/watch?v=kCc8FmEb1nY )

PartiallyTyped 3y ago

Attention is a convex linear combination of values (meaning the coefficients sum to 1) where the coefficients are generated at runtime. The values are given from the Value kernel; which is often just a linear map. The strength is given from the dot product of the query and the keys. The queries and keys are given from the Query and Key kernels. Both of which are usually linear maps. For each token, you have a key value query triplet (D,D,D). You take the query (D,1) and the matrix (N x D) of keys of all tokens, and do a Matrix vector product to get the scaling. You apply softmax to make it convex; and now you have a vector (N, 1). You do another matrix vector product with the Values mapping (N, D), which gives you the value for the current token (1,D). Repeat N times, and that’s it. If you would like to change which tokens each token considers, eg a token knows only of preceding tokens, you can do element wise product of the scaling vector with a mask vector which describes what you consider for current token ; and divide by their inner product. This will renormalise the valid coefficients and keep the rest 0; thus they are ignored when you combine all values together. If you would like to increase the number of attention heads; which can give you finer control of the scaling process, you can split the key query and value vectors into segments, one segment for each head. Eg first 2 indices go to head 1, indices 3,4 go to head 2, etc etc. then you use those to do attention, repeating the process above for each head; and then combine the results to a single vector. Idk if I confused query with keys, but this is the gist of it. In practice the operations are much more optimised than what I described.

soVeryTired 3y ago

Have you tried asking GPT-4? I got a pretty good explanation of attantion from it. The 'key' and 'value' description in the attention mechanism probably doesn't make much sense to you because at best it's a strained metaphor, rather than an actual principle by which attention operates. If you take a look at the maths, you can interchange the role of 'keys' and 'values' and get something mathematically identical.

metanonsense 3y ago

What made attention mechanism click to me was the video https://www.youtube.com/watch?v=0PjHri8tc1c https://www.youtube.com/watch?v=0PjHri8tc1c by Sebastian Raschka. Another good video on attention that tries to give a bit more intuition on the query/key/value stuff is https://www.youtube.com/watch?v=OyFJWRnt_AY https://www.youtube.com/watch?v=OyFJWRnt_AY In hindsight, what helped me most is understanding the dimensionality at each step of the attention computation (explained well in Raschka's video). Another think I simply did not get from the original Transformer paper is that the learning of self-attention happens in the linear layers. (At least, that's my current understanding. If that's wrong, I appreciate any correction :)

PartiallyTyped 3y ago
> Another think I simply did not get from the original Transformer paper is that the learning of self-attention happens in the linear layers. You can replace the KVQ kernels with any parametric computation that allows you to pass gradients through, and you will have learning. I think some newer language model architectures use residual blocks here, and some vision transformers use FeedForward networks for the kernels.
api 3y ago
Thanks. One thought I’ve had is to wonder if attention isn’t kind of like a way to unroll RNNs to make training more efficient. Any truth in that?
opyate 3y ago
I thoroughly recommend Sebastian's newsletter "Ahead Of AI". Link on his website: https://sebastianraschka.com/ https://sebastianraschka.com/

tomasff 3y ago

https://arxiv.org/abs/2207.09238 https://arxiv.org/abs/2207.09238 This is a great transparent and clear description of what is actually going on in attention layers. Alternatively, this article by Jay Alammar is also very popular http://jalammar.github.io/illustrated-transformer/ http://jalammar.github.io/illustrated-transformer/

opyate 3y ago
I commented yesterday, but came back to say this. Just found it, and it's really good. I like reading descriptions better than watching videos of them, so this is great.