Senthooran Rajamanoharan

Ω8110

UPDATE: we've corrected equations 9 and 10 in the paper (screenshot of the draft below) and also added a footnote that hopefully helps clarify the derivation. I've also attached a revised figure 6, showing that this doesn't change the overall story (for the mathematical reasons I mentioned in my previous comment). These will go up on arXiv, along with some other minor changes (like remembering to mention SAEs' widths), likely some point next week. Thanks again Sam for pointing this out!

Updated equations (draft):

Updated figure 6 (shrinkage comparison for GELU-1L):

Ω330

Yep, the intuition here indeed was that L1 penalised reconstruction seems to be okay for teaching a standard SAE's encoder to detect which features are on (even if features get shrunk as a result), so that is effectively what this auxiliary loss is teaching the gate sub-layer to do, alongside the sparsity penalty. (The key difference being we freeze the decoder in the auxiliary task, which the ablation study shows helps performance.) Maybe to put it another way, this was an auxiliary task that we had good evidence would teach the gate sublayer to detect active features reasonably well, and it turned out to give good results in practice. It's totally possible though that there are better auxiliary tasks (or even completely different loss functions) out there that we've not explored.

Ω110

Hey Sam, thanks - you're right. The definition of reconstruction bias is actually the argmin of

which I'd (incorrectly) rearranged as the expression in the paper. As a result, the optimum is

That being said, the derivation we gave was not quite right, as I'd incorrectly substituted the optimised loss rather than the original reconstruction loss, which makes equation (10) incorrect. However the difference between the two is small exactly when gamma is close to one (and indeed vanishes when there is no shrinkage), which is probably why we didn't pick this up. Anyway, we plan to correct these two equations and update the graphs, and will submit a revised version.

98

Thanks for sharing your findings - this was an interesting idea to test out! I played around with the notebook you linked to on this and noticed that the logistic regression *training* accuracy is also pretty low for earlier layers when using the encoded hidden representations. This was initially surprising (surely it should be easy to overfit with such a high dimensional input space and only ~1000 examples?) until I noticed that the number of 'on' features is pretty low, especially for early layer SAEs.

For example, the layer 2 SAE only has (the same) 2 features on over all examples in the dataset, so effectively you're training a classifier after doing a dimensionality reduction down to 2 dimensions. This may be a tall order even if you used (say) PCA to choose those 2 dimensions, but in the case of the pretrained SAE those two dimensions were chosen to optimise reconstruction on the full data distribution (of which this dataset is rather unrepresentative). The upshot is that unless you're lucky (and the SAE happened to pick features that correspond to sentiment), it makes sense you lose a lot of classification performance.

In contrast, the final SAEs have hundreds of features that are 'on' over the dataset, so even if none of those features directly relate to sentiment, the chances are good that you have preserved enough of the structure in the original hidden state to be able to recover sentiment. On the other hand, even at this end of the spectrum, note you haven't really projected to a higher dimensional space - you've gone from ~1000 dimensions to a similar or fewer number of effective dimensions - so it's not so surprising performance still doesn't match training a classifier on the hidden states directly.

All in all, I think this gave me a couple of useful insights:

- It's important to have really, really high fidelity with SAEs if you want to keep L0 (number of on features) low while at the same time be able to use the SAE for very narrow distribution analysis. (E.g. in this case, if the layer 2 SAE really had encoded the concept of sentiment, then it wouldn't have mattered that only 2 features were on on average across the dataset.)
- I originally shared your initial hypothesis (about projecting to a higher dimensional space making concepts more separable), but have updated to thinking that I shouldn't think of sparse "high dimensional" projections in the same way as dense projections. My new mental model for sparse projections is that you're actually projecting down to a
*lower*dimensional space, but where the projection is task dependent (i.e. the SAE's relu chooses which projections it thinks are relevant). (Think of it a bit like a mixture of experts dimensionality*reduction*algorithm.) So the act of projection will only help with classification performance if the dimensions chosen by the filter are actually relevant to the problem (which requires a really good SAE), otherwise you're likely to get worse performance than if you hadn't projected at all.

On bmag, it's unclear what a "natural" choice would be for setting this parameter in order to simplify the architecture further. One natural reference point is to set it to ermag⊙bgate, but this corresponds to getting rid of the discontinuity in the Jump ReLU (turning the magnitude encoder into a ReLU on multiplicatively rescaled gate encoder preactivations). Effectively (removing the now unnecessary auxiliary task), this would give results similar to the "baseline + rescale & shift" benchmark in section 5.2 of the paper, although probably worse, as we wouldn't have the shift.