Grounded: Inference via Local Signals & Learned Representations by Organic & Artificial Systems

Abstract

Before the brain was thought of as the seat of consciousness, it was believed to be a radiator to cool the humors that drove the body and mind. Before infant neurodevelopment was studied with electroencephalograms (EEGs) that examine the rhythms of periodic neuron activations, and before we sought to detect dementia early by analyzing the covariation of blood oxygen levels throughout the brain with functional magnetic resonance imaging (fMRI), it was thought you could know the intellectual capacity and character of a person by the shape of their skull. Through the work of numerous minds dedicating themselves to developing tools and theories, to gathering data and evidence that might help us better understand the brain, we have advanced our models beyond humors and phrenology. Over the last century, our understanding of neurons and the brain has progressed and inspired the creation of technology mimicking their action, which is now revolutionizing our interactions with each other and the world. This same technology is also enabling us even further insights into the nervous system. In the following work, I aim to explore substrates that produce the mind, more specifically the connections within networks of neurons (brain cells). Additionally, I will present efforts to extend the toolset available for this field of study, known as connectomics, to larger structures, larger portions of the brain than is currently feasible. The work herein contained, insofar as biology is concerned, regards the cerebellum of the mouse. Though strikingly regular in its structure, decades of examination have only continued to illuminate just how complex a computational apparatus the cerebellum truly is. It is to this tradition that this dissertation humbly contributes. Chapter 1 will proceed from a brief introduction to information, to some contrasts between natural and artificial algorithms for information processing, and finally to building a vocabulary for discussing the biological implementation of information processing in the cerebellum of the mouse. In chapter 2, I will present work examining how the specific connectivity statistics between mossy fibers and granule cells may enable noise-resilient pattern separation in the cerebellum of the mouse. In chapter 3, I will present findings supporting a preferential role of signals from the ascending portion of granule cell axons in guiding information processing by nearby Purkinje cells. In chapter 4, I will describe my application and interrogation of deep convolutional neural networks for the purposes of expanding our capacity for connectomic analyses of the brain. And finally, in chapter 5, for the inquisitive reader, I will further contextualize my work, discuss the deeper findings, and relate it to next steps and further fields of inquiry.