Associative learning in cortical and artificial neural networks

Abstract

One of the main goals of neuroscience is to explain the basic functions of the brain such as thought, learning, and control of movement. A comprehensive explanation of these functions must span different temporal and spatial scales to connect the workings of the brain at the molecular level to the circuit level to the level of behavior. This dissertation focuses on learning and formation of long-term memories - functions that are mediated by changes in synaptic connectivity. I examine the effects of learning on the connectivity and dynamics of networks in the brain and artificial neural networks. In the first chapter of this dissertation, I propose that many basic structural and dynamical properties of local cortical circuits result from associative learning. This hypothesis is tested in a network model of inhibitory and excitatory McCulloch and Pitts neurons loaded with associative sequences to capacity. I solve the learning problem analytically and numerically to show that such networks exhibit many ubiquities properties of local cortical citrus. These include structural properties, such as the probabilities of connections between inhibitory and excitatory neurons, distributions of weights for these connection types, overexpression of specific 2- and 3-neuron motifs, along with various properties of network dynamics. Because signal transmission in the brain is accompanied by many sources of errors and noise, in the second chapter of this dissertation I explore the effect of such unavoidable hindrances on learning and network properties. I argue that noise should not be viewed as a nuisance, but that it is an essential component of the reliable learning mechanism implemented by the brain. To test this hypothesis, I formulate and solve a biologically constrained network model of associative sequence learning in the presence of errors and noise. The results reveal that noise during learning increases the probability of memory retrieval and that it is required for optimal recovery of stored information. In the last chapter, I transition from biologically plausible artificial neuron network models of learning to a machine learning application. I develop a methodology for real-time automated reconstruction of neurons from 3D stacks of optical microscopy images. The pipeline is based on deep convolutional neural networks and includes image compression, image enhancement, segmentation of neuron cell bodies, and neurite tracing. I show that artificial neural networks can be trained to effectively compress 3D stacks of optical microscopy images and significantly enhance the intensity of neurites, making the results amenable for fast and accurate reconstruction of neurons.--Author's abstract