5 ms·

I think you're downplaying the importance of the attention/transformer architecture here. If it was "just" a matter of throwing compute at probabilities, then w

I think you're downplaying the importance of the attention/transformer architecture here. If it was "just" a matter of throwing compute at probabilities, then we wouldn't need any special architecture at all. P(next_word|previous_words) is ridiculously hard to estimate in a way that is actually useful. Remember how bad text generation used to be before GPT? There is innovation in discovering an architecture that makes it possible to learn P(next_word|previous_words), in addition to the computing techniques and hardware improvements required to make it work.

JeremyNT 3y ago
> There is innovation in discovering an architecture that makes it possible to learn P(next_word|previous_words), in addition to the computing techniques and hardware improvements required to make it work. Isn't that essentially what mjburgess said in the parent post? > LLMs are a hardware innovation: they make it possible for GPUs to compute this at scale across TBs of data... The algorithm isnt doing anything other than aligning computation to hardware
- nerdponx 3y ago
  Not really, and no. Torch and CUDA align computation to hardware. If it were just a matter of doing that, we would be fine with fully-connected MLP. And maybe that would work with orders of magnitude more data and compute than we currently throw at these models. But we are already pushing the cutting edge of those things to get useful results out of the specialized architecture. Choosing the right NN architecture is like feature engineering: the exact details don't matter that much, but getting the right overall structure can be the difference between learning a working model and failing to learn a working model, from the same source data with the same information content. Clearly our choice of inductive bias matters, and the transformer architecture is clearly an improvement over other designs. Surely you wouldn't argue that a CNN is "just" aligning computation to hardware, right? Transformers are clearly showing themselves as a reliably effective model architecture for text in the same way that CNNs are reliably effective for images.
  - mjburgess 3y ago
    Err... no. MLPs are fundamentally sequential algorithms (backprop weight updating). All major innovations in NN design have been to find ways of designing the architecture to fit GPU compute paradigms. It was an innovation, in the 80s, to map image structure to weight structure that underpins CNNs. That isnt what made CNNs trainable though.. that was alexnet, and just go read the paper... its pretty upfront about how the NN architecture is designed to fit the GPU... that's the point of it
  - rsfern 3y ago
    There’s some interesting work replacing scaled dot product attention and position embeddings with fixed format MLPs [0] - so I tend to lean towards thinking of classic transformers as having a reasonable enough inductive bias and the scalability to actually realize the amount of compute that’s needed 0: https://arxiv.org/abs/2105.08050 https://arxiv.org/abs/2105.08050
mjburgess 3y ago
Yes, it's really hard -- the innovation is aligning the really basic dot-product similarity mechanism to hardware. You can use basically any NN structure to do the same task, the issue is that they're untrainable because they arent parallizable. There is no innovation here in the sense of a brand new algorithm for modelling conditional probabilities -- the innovation is in adapting the algorithm for GPU training on text/etc.
- HarHarVeryFunny 3y ago
  > Yes, it's really hard -- the innovation is aligning the really basic dot-product similarity mechanism to hardware. You can use basically any NN structure to do the same task, the issue is that they're untrainable because they arent parallizable. This is only partially true. I wouldn't say you could use *any* NN architecture for sequence-to-sequence prediction. You either have to model them as a potentially infinite sequence with an RNN of some sort (e.g. LSTM), or, depending on the sequence type, model them as a hierarchy of sub-sequences, using something like a multi-layered convolution or transformer. The transformer is certainly well suited to current massively parallel hardware architectures, and this was also a large part of the motivation for the design. While the transformer isn't the only way to do seq-2-seq with neural nets, I think the reason it is so successful is more than simply being scalable and well matched to the available training hardware. Other techniques just don't work as well. From the mechanistic interpretability work that has been done so far, it seems that learnt "induction heads", utilizing the key-based attention, and layered architecture, are what give transformers their power.
- bruce343434 3y ago
  I don't know why you seem to have such a bone to pick with transformers but imo it's still interesting to learn about it, and reading your dismissively toned drivel of "just" and "simply" makes me tired. You're barking up the wrong tree man, what are you on about.
  - kordlessagain 3y ago
    Somebody's judgment weights need to be updated to include emoji embeddings.

JeremyNT 3y ago

> There is innovation in discovering an architecture that makes it possible to learn P(next_word|previous_words), in addition to the computing techniques and hardware improvements required to make it work. Isn't that essentially what mjburgess said in the parent post? > LLMs are a hardware innovation: they make it possible for GPUs to compute this at scale across TBs of data... The algorithm isnt doing anything other than aligning computation to hardware

nerdponx 3y ago
Not really, and no. Torch and CUDA align computation to hardware. If it were just a matter of doing that, we would be fine with fully-connected MLP. And maybe that would work with orders of magnitude more data and compute than we currently throw at these models. But we are already pushing the cutting edge of those things to get useful results out of the specialized architecture. Choosing the right NN architecture is like feature engineering: the exact details don't matter that much, but getting the right overall structure can be the difference between learning a working model and failing to learn a working model, from the same source data with the same information content. Clearly our choice of inductive bias matters, and the transformer architecture is clearly an improvement over other designs. Surely you wouldn't argue that a CNN is "just" aligning computation to hardware, right? Transformers are clearly showing themselves as a reliably effective model architecture for text in the same way that CNNs are reliably effective for images.
- mjburgess 3y ago
  Err... no. MLPs are fundamentally sequential algorithms (backprop weight updating). All major innovations in NN design have been to find ways of designing the architecture to fit GPU compute paradigms. It was an innovation, in the 80s, to map image structure to weight structure that underpins CNNs. That isnt what made CNNs trainable though.. that was alexnet, and just go read the paper... its pretty upfront about how the NN architecture is designed to fit the GPU... that's the point of it
- rsfern 3y ago
  There’s some interesting work replacing scaled dot product attention and position embeddings with fixed format MLPs [0] - so I tend to lean towards thinking of classic transformers as having a reasonable enough inductive bias and the scalability to actually realize the amount of compute that’s needed 0: https://arxiv.org/abs/2105.08050 https://arxiv.org/abs/2105.08050

mjburgess 3y ago

Yes, it's really hard -- the innovation is aligning the really basic dot-product similarity mechanism to hardware. You can use basically any NN structure to do the same task, the issue is that they're untrainable because they arent parallizable. There is no innovation here in the sense of a brand new algorithm for modelling conditional probabilities -- the innovation is in adapting the algorithm for GPU training on text/etc.

HarHarVeryFunny 3y ago
> Yes, it's really hard -- the innovation is aligning the really basic dot-product similarity mechanism to hardware. You can use basically any NN structure to do the same task, the issue is that they're untrainable because they arent parallizable. This is only partially true. I wouldn't say you could use *any* NN architecture for sequence-to-sequence prediction. You either have to model them as a potentially infinite sequence with an RNN of some sort (e.g. LSTM), or, depending on the sequence type, model them as a hierarchy of sub-sequences, using something like a multi-layered convolution or transformer. The transformer is certainly well suited to current massively parallel hardware architectures, and this was also a large part of the motivation for the design. While the transformer isn't the only way to do seq-2-seq with neural nets, I think the reason it is so successful is more than simply being scalable and well matched to the available training hardware. Other techniques just don't work as well. From the mechanistic interpretability work that has been done so far, it seems that learnt "induction heads", utilizing the key-based attention, and layered architecture, are what give transformers their power.
bruce343434 3y ago
I don't know why you seem to have such a bone to pick with transformers but imo it's still interesting to learn about it, and reading your dismissively toned drivel of "just" and "simply" makes me tired. You're barking up the wrong tree man, what are you on about.
- kordlessagain 3y ago
  Somebody's judgment weights need to be updated to include emoji embeddings.