Summary

I sandwich the sparse layer in a sparse autoencoder (SAE) between non-sparse lower-dimensional layers and refer to this as a deep SAE.
I find that features from deep SAEs are at least as interpretable as features from standard shallow SAEs.
I claim that this is not a tremendously likely result if you assume that the success of SAEs is entirely explained by the accuracy of the superposition hypothesis.
I speculate that perhaps by relaxing our adherence to the concrete principles laid out by the superposition hypothesis, we could improve SAEs in new ways.

Introduction

Context

Instead of rehashing the superposition hypothesis and SAEs, I will just link these wonderful resources: Toy Models of Superposition and Towards Monosemanticity.

Motivations

My sense is that the standard justification for the success of SAEs is that the superposition hypothesis is just an accurate model of neural network function, which directly translates to SAE effectiveness. While many of the arguments for the superposition hypothesis are compelling, I had a growing sense that the superposition hypothesis was at least not a sufficient explanation for the success of SAEs.

This sense came from, for example, inconsistencies between empirical findings in the posts linked above, for example that features will take similar directions in activation space if they don’t co-occur often, but also if they have similar effects on downstream model outputs, which seems contradictory in many cases. I won't elaborate too much on this intuition because (1) I don’t think understanding this intuition is actually that necessary to appreciate these results even if it served to initiate the project, (2) in hindsight, I don’t even think that these results strongly confirm my original intuition (but they also don’t oppose it).

Ultimately, I posited that, if there is some unidentified reason for the success of SAEs, it might be that sparsity is just a property of representations that us humans prefer, in some more abstract sense. If this were true, we should directly aim our SAEs to produce the most faithful and sparse representations of neural network function as possible, possibly abandoning some of the concrete principles laid out by the superposition hypothesis. And it seems that the obvious way to do this is to add more layers.

I'm posting my work so far because:

I want to see if I'm missing something obvi

...

adam-oberman

adam-oberman