Beyond $\ell_1$ sparse coding in V1

TL;DR

This paper demonstrates that biological visual cortex likely employs a sparsity regularization closer to the $ ext{l}_0$ pseudo-norm rather than the traditional $ ext{l}_1$ norm, leading to more efficient and biologically plausible sparse coding.

Contribution

The study shows that $ ext{l}_1$ sparsity is suboptimal for modeling V1 and proposes that V1 uses a regularization closer to $ ext{l}_0$, improving biological plausibility and feature representation.

Findings

$ ext{l}_1$ sparsity produces denser codes and higher overcompleteness.

Methods closer to $ ext{l}_0$ yield more V1-like receptive fields.

Soft thresholding degrades reconstruction performance compared to $ ext{l}_0$-like regularization.

Abstract

Growing evidence indicates that only a sparse subset from a pool of sensory neurons is active for the encoding of visual stimuli at any instant in time. Traditionally, to replicate such biological sparsity, generative models have been using the $\ell_1$ norm as a penalty due to its convexity, which makes it amenable to fast and simple algorithmic solvers. In this work, we use biological vision as a test-bed and show that the soft thresholding operation associated to the use of the $\ell_1$ norm is highly suboptimal compared to other functions suited to approximating $\ell_q$ with $0 \leq q < 1 $ (including recently proposed Continuous Exact relaxations), both in terms of performance and in the production of features that are akin to signatures of the primary visual cortex. We show that $\ell_1$ sparsity produces a denser code or employs a pool with more neurons, i.e. has a higher degree of overcompleteness, in order to maintain the same reconstruction error as the other methods considered. For all the penalty functions tested, a subset of the neurons develop orientation selectivity similarly to V1 neurons. When their code is sparse enough, the methods also develop receptive fields with varying functionalities, another signature of V1. Compared to other methods, soft thresholding achieves this level of sparsity at the expense of much degraded reconstruction performance, that more likely than not is not acceptable in biological vision. Our results indicate that V1 uses a sparsity inducing regularization that is closer to the $\ell_0$ pseudo-norm rather than to the $\ell_1$ norm.