Network Modelling and Functional Graph Structure of the Primary Visual Cortex

主視覺皮層的網絡建模和功能圖結構

Student thesis: Doctoral Thesis

View graph of relations

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date23 Sep 2021

Abstract

Biological vision is a fast, flexible, and energy-efficient sensing system that mammals utilize to perceive the world. Understanding how biological vision systems achieve their characteristic combination of speed, flexibility, and energy-efficiency can pro­vide insights for improving computer vision systems and assistive technologies. One limiting factor of this line of research is a bias in visual neuroscience literature towards neuronal-­level descriptions of computation that obfuscate the network­-level computa­tional structure of the mammalian visual system. No neuron functions in isolation: neu­ral networks code information through coordinated behavior. Populations of neurons code information in ways that are not discernible from the activity of single neurons. Calcium imaging can now simultaneously record larger populations than electrophysi­ological techniques could previously, allowing single-neuron analyses to be expanded to the network level. Larger datasets demand the development of computational ap­proaches for network-level analysis. Graph theory, which conceptualizes networks as series of nodes interacting via edges, can efficiently characterize network types and elucidate important network properties. However, for the foreseeable future there still exists a subsampling problem in neuroscience experimentation. There are more than70million neurons in the mouse brain, but only hundreds of neurons can be recorded simultaneously. This subsampling problem motivates development of computational techniques that can estimate the unobserved inputs driving the behavior of observ­able neurons in the brain, such inferred inputs to the observed system are called latent variables(LVs). To extract these latent variables driving the behavior of a neuronal network, specific latent variable models have been developed. Latent variable mod­els constrained by known properties of neuronal networks, such as non-­negativity and weak correlations, are excellent tools for inferring unobserved brain function. Taken together, the need for a network-level understanding of biological visual systems and the experimental limitations of small sample sizes, positions graph theory and latent variable models in synergistic roles for advancing computational neuroscience. This thesis utilized a large dataset, made publicly available by the Allen Brain Institute. Recordings from the primary visual cortex of the mouse were used to characterize the network properties of, and extract meaningful latent features from, the biological vision system.

One major driving force in the rapid advancement of machine learning in the last decade was the development of Deep Convolutional Neural Networks (DCNNs). This network architecture was initially inspired by the mammalian visual system structure, mimicking the hierarchical feed-­forward architecture in the low-levels of the mam­malian visual system. DCNNs such as VGG16, ImageNet, and ResNet, among others, have been successful at object recognition tasks far beyond the capabilities of tradi­tional computer vision systems. The original biologically­-inspired design of DCNNs, and their successful application in the machine learning field, has sparked interest in the neuroscience community that DCNNs may be helpful for studying the mammalian 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 system and artificial neural network structure is an open ques­tion. Using graph­-theoretic measures, this thesis compared the biological system with a modern computer vision system (VGG16), showing consistent differences between the biological and artificial systems, and leading to the hypothesis that the artificial system would perform more efficiently if training strategies transformed the network node density distribution from Gaussian toward Poisson.

This thesis describes an efficient graph theory-based method for model selection that was developed from the observation that node density is able to predict the recon­struction performance of two types of non-­negative latent variable models. Along with reconstruction performance, the unobserved drivers of neuronal behavior extracted by these latent variable models were evaluated herein. These extracted latent variables illustrate the complexity of neuronal behavior profiles in response to a simple visual stimulus. This analysis shows that the typical two-dimensional representation of neu­ronal response to drifting gratings stimulus is an over­-simplification of the behavior of primary visual cortex neurons. I show two more important findings that elucidate the behavior of latent variable models in the neuroscientific context: I demonstrated that it is not always necessary to deconvolve calcium fluorescence traces into action poten­tials in order to fit the latent variable models to neuronal data, and I demonstrated that the latent variables extracted via these modelling techniques do not necessarily result in more simple representations of the stimulus, but retain the diversity and complexity of neuronal behavioral profiles.

While the majority of my studies focused on murine primary visual cortex, I also extended my research to investigate image feature effects on human perceptions. To better understand which image features could communicate complex emotional infor­mation to humans, I have conducted an experiment with collaborators to investigate human emotional response to images. The results show that the significant image fea­tures for human emotional response are mostly semantic features, and more complex than previously reported. This preliminary experiment underpins ongoing work in our laboratory including a pending USA patent application for a computer-vision-based emotion recognition and style transfer system.

The following thesis elucidates insights gained from graph theoretic description, and latent variable modelling, of the murine primary visual cortex. These insights gen­erated further testable hypotheses that have already proven useful in the development of computer vision and assistive technologies.

    Research areas

  • Vision, Convolutional Neural Networks, Latent Variable Models, Non-Negative Matrix Factorization, Rectified Latent Variable Model, Calcium Imaging, Neural networks (Neurobiology), Neural networks (Computer science), Emotion Estimation