Comparing the brain's representation of shape to that of a deep convolutional neural network

Abstract

Hierarchical neural nets are currently the highest performing general purpose image recognition computer algorithms. Their design is loosely inspired by the neural architecture of the ventral visual pathway in the primate brain, which is believed to underlie the perception of form and the ability to recognize objects. The exact tuning of artificial neural units within an HNN, however, is not prescribed from known biology, but is trained using a performance-based learning algorithm. In evaluating whether HNNs are ripe for further bio-inspired performance improvements, it is of interest to test whether the response properties in the intermediate layers of the HNN approximate those of the ventral visual stream. We therefore characterized units within a popular HNN with a set of visual stimuli that has been employed by neurophysiologists to successfully characterize the shape-tuning properties of neurons in the intermediate visual cortical area V4 of the ventral stream. We found that the tuning and fits of a small but significant number of units in the HNN were strikingly similar to those of some V4 neurons for our simple set of test shapes. There tended to be more such units in the deeper layers of the HNN. We discuss implications of our results to the encoding of curvature in the primate brain and propose ways to further characterize V4-like shape tuning in HNNs.