Comparing Biological and Artificial Neural Networks: Effects of Sampling on the Quantification of Network Properties

Description

The mammalian neural system continues to evade a full quantitative characterization. Advances in Deep Convolutional Neural Networks (DCNN) have provided new opportunities for computational neuroscience to pose new questions regarding the structure and function of biological neural networks, particularly the visual system. Some attempts have been made to utilize advances in machine learning to answer neuroscientific questions, but how to appropriately make comparisons between the biological neural systems and DCNN models is an open question.To better understand biological neural network, neural recordings during animal behavior are necessary. However, current experimental techniques impose spatial limits on the number of neuronal units that can be recorded. To model neural dynamics in response to stimulus, various mathematical models have been developed to link external stimuli to the electrical or chemical activities of neurons. Yet, difficulty in performing in vivo experiments result in the limited quantity of training data, which could still severely affect the accuracy of the modeling of the biological neural network, as the neural spiking activity is already quite sparse. The spatiotemporal limitations in recordings also affect the reconstruction of networks, which describe the connections between the neuronal units recorded.In order to investigate the similarities and differences between network properties of biological nervous system and artificial neural networks, it is necessary to first investigate how sampling affects complex network measures. We will focus on the mouse visual system and different DCNN architectures for machine vision, to determine if this comparison is appropriate. Utilizing weighted graph-theoretic measures, we will investigate the connectivity patterns in the modern artificial computer vision system and the biological vision system. We will then extend the work to investigate the network properties of other brain regions and artificial neural networks for unique functions, such as motor and auditory functions, and to examine how different neural network structures relate to function and/or performance. A better understanding of how information is processed in biological neural circuits will in turn contribute to the development of artificial neural networks and neuromorphic circuits.