Neural networks as models of the brain : insights on biological learning rules, limiting dynamics of training and model-brain comparison
Abstract/Contents
- Abstract
- This dissertation explores the intersection of Neuroscience and Artificial Intelligence, focusing on the mutual understanding that can be derived between the study of the brain and the development of Artificial Neural Networks (ANNs). The models studied in this work are derived in the context of goal-driven modeling, which describes ANNs by specifying four components: the objective function to optimize, the dataset used, the architecture of the ANN, and the learning rule used to iteratively update its parameters. The first chapters of this dissertation focus on the learning rule component. In Chapter 2, we explore the biological plausibility of backpropagation and propose performant learning rules that could be implemented in the brain while performing approximate backpropagation. Chapter 3 theoretically studies the limiting dynamics of weights in ANNs trained with SGD using stochastic differential equations and draws connections between the hyperparameters of the learning algorithm and the characteristics of the limiting solution. Still, if we build a goal-driven model of the brain, a crucial question remains: How can we effectively evaluate it as an accurate representation of the brain? Existing approaches based on behavioral decoding or comparing population-level representations are informative but have limitations. In the final chapter, we will argue for an approach based on single-unit response predictivity using a generalized linear model (GLM). We arrive at this particular GLM following a principled approach that seeks to predict brain responses from ANN responses using the same class of functions one would use to predict one specimen's brain responses from another specimen's. The ideal function class should exhibit high predictive accuracy when mapping between equivalent brain areas in a pair of specimens and low predictive accuracy when mapping between different brain areas. To measure these criteria, we define similarity and separability scores. We simulate a population of models of the mouse visual system and evaluate a variety of function classes on these two metrics. By comparing our results to neural data from mouse visual cortex electrophysiology, we refine the candidate function class and provide preliminary evidence that a single-unit predictivity metric based on this function class could better differentiate brain models.
Description
Type of resource | text |
---|---|
Form | electronic resource; remote; computer; online resource |
Extent | 1 online resource. |
Place | California |
Place | [Stanford, California] |
Publisher | [Stanford University] |
Copyright date | 2023; ©2023 |
Publication date | 2023; 2023 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | Sagastuy Brena, Javier |
---|---|
Degree supervisor | Yamins, Daniel |
Thesis advisor | Yamins, Daniel |
Thesis advisor | Ganguli, Surya, 1977- |
Thesis advisor | Hastie, Trevor |
Thesis advisor | Linderman, Scott |
Degree committee member | Ganguli, Surya, 1977- |
Degree committee member | Hastie, Trevor |
Degree committee member | Linderman, Scott |
Associated with | Stanford University, School of Engineering |
Associated with | Stanford University, Institute for Computational and Mathematical Engineering |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Javier Sagastuy-Brena. |
---|---|
Note | Submitted to the Institute for Computational and Mathematical Engineering. |
Thesis | Thesis Ph.D. Stanford University 2023. |
Location | https://purl.stanford.edu/qp892gq0847 |
Access conditions
- Copyright
- © 2023 by Javier Sagastuy Brena
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...