5 ms·
Took me a while to understand what their "hot" and "cold" neurons meant, since in most ML I do, there is no such notion. And their paper doesn't directly defin
by phh 3y ago
Took me a while to understand what their "hot" and "cold" neurons meant, since in most ML I do, there is no such notion. And their paper doesn't directly define it (or I missed it)
After some thoughts, in ReLU it does make sense, because half of the function is constant, so you can say that you're "cold" if that neuron's ReLU-ed output is often 0 . So I checked whether ReLU was common in LLMs, original llama doesn't use ReLU. But after (re-)reading the github, it actually only works on ReLU models. Turns out that there is a group of people "fine-tuning" (I would rather call that re-training, since you start by breaking the model?) models to use ReLU to allow for that sparsity: https://huggingface.co/SparseLLM https://huggingface.co/SparseLLM
So this is sadly not applicable to any model you can find on the internet, but that sounds like a great progress anyway. Possibly this might shift the compromises back to bigger models but with "less ideal" activations. Also I'm curious what would be the legal impacts on it (since USA and EU refers to a model's FLOPs/number of parameters... How do you compute it with sparsity? Do you average?)
I think that a possible avenue for future research in that area is keeping original activation (like llama keeping SwiGLU), but using quantification to define "hot" and "cold" neurons to be saturation areas. (For example, saying that this activation function, below -1. at 8 bit, is equivalent to -infinity, and thus this is a cold neuron)
- brucethemoose2 3y agoThat is a huge caveat to leave out of a readme, especially one that claims llama compatibility.
- ShamelessC 3y agoThey don’t make that claim as far as I can tell. Just that they support llama2 models.
- brucethemoose2 3y agoWell it's not really support of llama 2 if it has to be extensively finetuned to "convert" the model.
- acqq 3y agoIndeed https://huggingface.co/SparseLLM/ReluFalcon-40B https://huggingface.co/SparseLLM/ReluFalcon-40B "We utilize PowerInfer for inference"
- boredumb 3y ago> Also I'm curious what would be the legal impacts on it (since USA and EU refers to a model's FLOPs/number of parameters... How do you compute it with sparsity? Do you average?). How/when did these types of regulations come about? This feels like an insane thing to have to keep in mind while developing.
- radicalbyte 3y agoThe EU messed up with the GDPR - they should have implemented it at least a decade earlier and ignored the lobby which lead to the cookie banner instead of either an outright ban on tracking for all but a tiny number of purposes. Such a ban would have had a negligible impact on the tech industry financially but would have had huge privacy rewards. They're trying to get in early on AI so as not to make the same mistake again. Which might result in them making the opposite mistake.
- quocanh 3y agoTiny negligible impact on the industry (Except cut advertising revenue in half, but who cares. What do ads pay for anyways?)
- Nextgrid 3y ago> What do ads pay for anyways? Making the world a worse place? If you look carefully you’ll realize most of the harms and negative effects of technology are due to it being primarily funded by advertising and trying to maximize ad revenue.