Ali Almasi
University of Melbourne and & National
Vision Research Institute, Australia
Title: Feature selectivity and invariance in primary visual cortex
Abstract:
Object
recognition in scenes develops across a hierarchy of visual areas. Robust
recognition requires fine selectivity for particular features
of relevance and invariance to irrelevant features. Deep convolutional neural
networks have achieved near-human levels of performance in object recognition
by iteratively applying filters that select features, followed by pooling of
their outputs to generate invariance. We applied a “filter-then-pool” model to
recordings from neurons in cat primary visual cortex (V1) to investigate visual
feature selectivity and invariance in the brain. Many neurons pooled the
outputs of multiple filters, resulting in selectivity for feature
characteristics that were preserved across filters, and invariance to feature
characteristics that differed across filters. We found cells corresponding to
the “energy model” of V1 complex cells that were invariant to spatial phase but
selective to a combination of other feature characteristics. We also frequently
found cells that showed only partial invariance to spatial phase, while
exhibiting invariance to perturbations in peak orientation and spatial
frequency. For each of these feature characteristics, some cells were more
selective for the characteristic and others that were more invariant. To
quantify the “selective-to-invariant” spectrum we used a bandwidth measuring
the range of a characteristic over which a cell responded equally allowing for
modest changes in contrast (< 2x). Peak orientation had the greatest portion
of selective cells, followed by peak spatial frequency and then spatial phase,
the latter showing the greatest portion of invariance. The bandwidth measure
also allowed us to quantify how much the nonlinear pooling operation
contributed to invariance by comparing it to the bandwidth expected from a
linear operation. This showed that spatial phase invariance benefited the most
from nonlinear pooling over multiple features, with the bandwidth frequently
much greater than expected in the linear case, and sometimes reaching the maximum
possible (360 deg). In contrast, orientation and spatial frequency had
bandwidths that did not increase much with nonlinear pooling over that of
linear pooling. Thus, in V1 there is a diversity of cells that combine
selectivity for some feature characteristics with invariance to perturbations
in others. This diversity encompasses a variety of feature characteristics
beyond spatial phase.
Bio-sketch:
Ali completed his PhD at the university of Melbourne in 2017 and is currently a research fellow at the National Vision Research Institute of Australia. His research interests include understanding how visual information is processed in the brain using experimental and theoretical approaches. During his PhD, Ali investigated the applicability of computational models based on efficient coding to complex cells in primary visual cortex, which is challenging topic due to the inherent nonlinearities found in their receptive field properties.