softmax should be exp()/1+∑exp()
Notice the 1 added to the denominator.
The difference is at the negative limit, softmax can be 0, instead of some epsilon. The same could be done by adding an extra zero value in x.
Downside is, you have to retrain your model from scratch to fix this.
if you think about it, the "escape hatch" is the design of the entire transformer dictionary. if Key/Query attention misaligns with Value's weights, you get a layer head that does not attend to anything...
Still, differential attention is pretty interesting & the benchmarking good, seems worth a try! It's in the same vein as linear or non-softmax attention, which also can work.
Note that there is an error below Eq. 1: W^V should be shape [d_model x d_model] not [d_model, 2*d_model] as in the Q, K matrices.
Idea: why not replace the lambda parameterization between softmax operations with something more general, like a matrix or MLP? E.g: Attention is the affine combination of N softmax attention operations (say, across heads). If the transformer learns an identity matrix here, then you know the original formulation was correct for the data; if it's sparse, these guys were right; if it's something else entirely then who knows...
I didn't follow Miller's proposal quite as he wrote it though and I put the mechanism in all the layers rather than avoiding it at the end.
My test doesn't absolutely rule out usefulness-- there's always different ways of applying something, but I saw no indication of it.
A/B test the two models and compare?
Would be interesting to see if these activations only show up on larger models, or they're some relation to model size.
Hah. Yes. It looks like they only show up in models with 6.7B parameters or more.
The problem can start at 125M. Small enough to test on a whim.
So train a model that exhibits these behaviours, then try it out.
> softmax should be exp()/1+∑exp()
I've also tested it on many networks and different types of attention and I've yet to see a meaningful improvement (or even an improvement), even in generalization.
It really is the training method...
As to the paper, I'm also still at a big lost and honestly, if reviewing could not accept it. The results look good, but I can't tell why and there's some "black magic" going on here.
- Figure 3 has "Transformer" and doesn't specify. Is this StableLM-3B-4E1T?
- What fucking dataset is this on? Stable has a WandB link[2] for that project and I don't see any experiment with similar (presumably entropy?) loss values (come on... this is fucking research... label your fucking graphs...)
- Where the fuck is the ablation? (Yes, I saw Fig 6 and Sec 3.8)
- How do I know that (assuming this is Stable) that the difference isn't just hyperparemeters? Or worse, GPUs! (yes, number of GPUs can change results due to sharding and this changing the statistics)
- How do I know it isn't down to 1k warmup steps instead of 5k?
- What about hidden size, layers, heads, or FFN size? Stable has 32/2560/32/? and this has 28/3072/12/8192 (these all will mess with sharding statistics too). Is the head dimension the same?
- How do I know it isn't down to the tokenizer?
- What is this magic? `0.8 - 0.6 * math.exp(-0.3 * depth)`
- Was this learned? Hand picked? This is a huge factor
- Any information about the learned parameters? Their final values? Trajectories?
- The code does not seem to be the same as whats in the algos...
[0] https://www.evanmiller.org/attention-is-off-by-one.html
[1] This is a bit convoluted but without this condition many "alternative forms" you see would be equivalent to other architectures like linear layers or gated units. Term is not well defined, but this really appears to be the only agreed upon aspect, even if only implicitly stated. This is a much longer conversation though.
[2] https://stability.wandb.io/stability-llm/stable-lm/reports/StableLM-3B-4E1T--VmlldzoyMjU4?accessToken=u3zujipenkx5g7rtcj9qojjgxpconyjktjkli2po09nffrffdhhchq045vp0wyfo
[2.1] The config: https://github.com/Stability-AI/StableLM/blob/main/configs/stablelm-3b-4e1t.ymlI know he's an economist btw. I was also surprised he got a job at anthropic a few months after. I wonder if they're related.
Take a vanilla MHA, tie the V projection between consecutive heads, make the output projection subtract consecutive heads, with some fixed prefactor and voila, you're most if not all of the way there.