Reevaluating Sparse Autoencoders: The Dictionary Dilemma

For years, the linear representation hypothesis has shaped how we see neural network activations. It's all about encoding high-level concepts as linear mixtures. But what happens when these mixtures, under superposition, project from a vast concept space into a cramped activation space? The theory suggests that a linear line in the concept world might twist and turn once it's projected. Sparse autoencoders (SAEs) have attempted to tackle this, but there's a glaring issue.

The Amortization Myth

SAEs try to compress sparse inference into a static encoder. The catch? This system falls short when faced with out-of-distribution (OOD) compositional shifts. Controlled experiments reveal that this isn't just a fluke. It's a persistent gap that haunts SAEs across diverse training set sizes, latent dimensions, and sparsity levels.

So, what's the real bottleneck? Turns out, it's not the inference procedure. It's the dictionary learning. SAEs are armed with dictionaries that point in the wrong directions. Even if you swap the encoder for per-sample Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) using the same faulty dictionary, the gap remains.

Scalable Dictionary Learning: The True Challenge

What does this mean for the future of sparse inference? It's time to shift our focus from amortization to scalable dictionary learning. An oracle baseline demonstrates that with the right dictionary, we can solve the problem at any scale. It's a bold claim, but the evidence is mounting.

Why should anyone in AI care? Because slapping a model on a GPU rental isn't a convergence thesis. We need to build systems that can adapt to shifts in data distribution. If SAEs can't handle these shifts due to flawed dictionaries, it's a wake-up call for researchers and practitioners alike.

The question remains: will the community rise to the occasion and redefine dictionary learning for sparse inference? Or will we continue to pour resources into a flawed system?