Increasing neural network robustness improves match to macaque V1 eigenspectrum, spatial frequency preference and predictivity

Abstract

Task-optimized convolutional neural networks (CNNs) show striking similarities to the ventral visual stream. However, human-imperceptible image perturbations can cause a CNN to make incorrect predictions. Here we provide insight into this brittleness by investigating the representations of models that are either robust or not robust to image perturbations. Theory suggests that the robustness of a system to these perturbations could be related to the power law exponent of the eigenspectrum of its set of neural responses, where power law exponents closer to and larger than one would indicate a system that is less susceptible to input perturbations. We show that neural responses in mouse and macaque primary visual cortex (V1) obey the predictions of this theory, where their eigenspectra have power law exponents of at least one. We also find that the eigenspectra of model representations decay slowly relative to those observed in neurophysiology and that robust models have eigenspectra that decay slightly faster and have higher power law exponents than those of non-robust models. The slow decay of the eigenspectra suggests that substantial variance in the model responses is related to the encoding of fine stimulus features. We therefore investigated the spatial frequency tuning of artificial neurons and found that a large proportion of them preferred high spatial frequencies and that robust models had preferred spatial frequency distributions more aligned with the measured spatial frequency distribution of macaque V1 cells. Furthermore, robust models were quantitatively better models of V1 than non-robust models. Our results are consistent with other findings that there is a misalignment between human and machine perception. They also suggest that it may be useful to penalize slow-decaying eigenspectra or to bias models to extract features of lower spatial frequencies during task-optimization in order to improve robustness and V1 neural response predictivity.

Anatomical and neurophysiological data on primary visual cortex suffice for reproducing brain-like robust multiplexing of visual function

“…We have also reproduced the result of [Kong et al, 2022] that feedforward CNN has a substantially smaller value of a (Fig. 4D) and found that the recurrent CNN model of [Wang and Hu, 2021], which also achieves very high accuracy for image classification, has an even smaller value of a (Fig.…”

“…In-vivo recordings of [Stringer et al, 2019] found that the value of α for primary visual cortex of mouse is actually close to this optimal value α = 1 + 2/ d . This neural coding property of area V1 in the brain has gained additional interest through the contrasting result of [Kong et al, 2022]. They found that feedforward CNNs, which are viewed to be substantially less noise robust than the brain, have in fact a smaller a, as predicted by the theory.…”

“…But the neural representation in the model of Billeh et al is more robust because it employs, like the brain [Stringer et al, 2019], a power law for the explained variance in higher PCA components that is close to a theoretically optimal compromise between the opposing goals to create noise-robust neural codes and to capture many details of visual inputs. In contrast, neural codes in CNNs have a different power law which reveals a focus on the latter [Kong et al, 2022]. In addition, we demonstrate that the model of Billeh et al uses preferentially those dimensions of population activity for coding that are orthogonal to the largest noise dimensions, like the brain does [Rumyantsev et al, 2020].…”

Not so spontaneous: Multi-dimensional representations of behaviors and context in sensory areas

“…Like SA, sensory stimuli also activate a diverse set of neural patterns, and various studies have observed tens to hundreds of significant dimensions in flies, zebrafish, mice, and primates (Avitan et al, 2017;Cowley et al, 2016;Kong et al, 2022;Murthy et al, 2008;Stringer et al, 2019a). Many neurons are active during both spontaneous and sensory-driven activity, but it appears that the neurons present in a specific SA pattern are not all driven by the same sensory stimulus.…”

Investigating Power laws in Deep Representation Learning

“…the variance explained by the n th principal component of the covariance matrix scaled as 1/n. In recent work (Kong et al, 2022) observe similar structure in representations from V1 in macaque monkeys and relate robustness to eigenspectrums of neural activations. Put another way, the coefficient of the power law (often denoted α), has been observed close to 1 in biological brains.…”