Comparing biological and artificial vision systems: Network measures of functional connectivity

Abstract

Advances in Deep Convolutional Neural Networks (DCNN) provide new opportunities for computational neuroscience to pose novel questions regarding the function of biological visual systems. Some attempts have been made to utilize advances in machine learning to answer neuroscientific questions, but how to appropriately make comparisons between the biological systems and artificial neural network structure is an open question. This analysis quantifies network properties of the mouse visual system and a common DCNN model (VGG16), to determine if this comparison is appropriate. Utilizing weighted graph-theoretic measures of node density (weighted node-degree), path length, local clustering coefficient, and betweenness, differences in functional connectivity patterns in the modern artificial computer vision system and the biological vision system are quantified. Results show that the mouse exhibits network measure distributions more similar to Poisson than normal, whereas the VGG16 exhibits network measure distributions with a more Gaussian shape than the sampled biological network. The artificial network shows higher density measures and shorter path lengths in comparison to the biological network. These results show that training a VGG16 for an object recognition task is unlikely to produce a network whose functional connectivity is similar to the mammalian visual system.