| dc.description.abstract | Neuroscience and artificial intelligence (AI) research have long enjoyed a synergistic relationship. AI has drawn key inspiration from the organization and function of the brain, while our understanding of the biological processes underlying computation has been profoundly enriched by studying the behavior of artificial systems. As breakthroughs in generative AI continue to transform our world, and as the need for more sustainable artificial neural systems becomes more urgent, the neuro-AI feedback loop has never been more important. AI needs ever more powerful and efficient systems, and neuroscience needs further insight into how our brains work. The development of more brain-like AI promises solutions to both of these problems. Unfortunately, this has thus far been stymied by two critical challenges: 1) how do we identify the features that make a system brain-like and 2) how do we incorporate these features into artificial networks in a useful and interpretable way? To address the first of these challenges, I will use the remarkable structural and biophysical diversity of the brain as an introduction into what it means for a system to be “brain-like.” This will lead us to a discussion of dendrites, the tree-like structures implicated at virtually every length scale of neural computation. Dendrites will thereafter act as the focal point for our study of brain-like computation. Specifically, I will trace how relatively simple biophysical features defined at the subcellular level can transform the computational landscape of large networks of neurons. To address the second of these challenges, it is necessary to discuss several enduring problems in computational neuroscience, broken down as chapters in this thesis. In Chapter 2, I will present the development of a new model of single-neuron dynamics that is realistic enough to capture the rich dynamics of dendritic spiking but efficient enough for use in simulations of thousands of neurons, thereby filling a long unmet need in the field. In Chapter 3, I will describe a solution to the general problem of training neural networks with arbitrary differentiable dynamics, thus opening the door for the study of countless biophysical phenomena in the context of networks that can learn to perform computations. In Chapter 4, I will use these tools to test several longstanding hypotheses regarding the utility of different biophysical features in neurons, performing first-of-their-kind fair comparisons of the computational performance of spiking networks, rate-based networks, and networks with nonlinear and linear dendrites. Finally, in Chapter 5, I will use insights gained from studying dendrites at the network level to provide a new perspective as to how the structural and biophysical diversity of the brain could emerge from a complex interplay of functional pressures (e.g., task demands) and physical constraints (e.g., space and energy). Together, the chapters of this thesis outline a general quantitative framework for building more brain-like AI for use in both AI research and neuroscience. This framework illustrates how biophysical specializations arising at the level of single neurons shape the emergent dynamics of the brain. | |
