Neural computation through the lense of recurrent dynamics
Abstract/Contents
- Abstract
- Recurrent connections, consisting of not only feedforward but also feedback connections between neurons, shape dynamics of population neural activity, which is thought to underlie our mental processes. Despite their ubiquity in the brain, their computational roles are less understood compared to their feedforward counterpart. In this dissertation, we present three studies that make progress toward improving our understanding of the recurrent neural dynamics. In the first part of the dissertation, we address the question "how does the brain perform computations while being robust to perturbations?" in the context of a short-term memory task, which relies on recurrent neural dynamics. We found that a frontal cortical region called anterior lateral motor cortex (ALM) that maintains short-term memory in this task achieves robustness through a modular organization which prevents spreading of a perturbation on one hemisphere to the other hemisphere. In addition, this region exhibited an error-correction property whereby the short-term memory of a perturbed hemisphere is rapidly recovered by input from the other intact hemisphere. We present evidence that a state-dependent gating of interhemispheric interactions underlies both the modularity and error correction property of ALM, through dynamic modeling of single-trial neural data and optogenetic perturbation experiments. Furthermore, we studied how such robust dynamics could arise in a highly connected network of neurons through recurrent neural network (RNN) modeling, and identified three necessary and sufficient conditions that made RNNs robust. This provides insight into how ALM can develop robust dynamics without experiencing perturbations during learning. In the second part of the dissertation, we tackle the problem of identifying the most "useful" perturbations to dissect a given neural circuit. We propose that if a perturbation is to be useful, it should be able to maximally discriminate distinct hypotheses about a neural circuit, and provide a mathematical framework to derive "optimal" perturbations according to this criterion. Specifically, we formulate distinct hypotheses as distinct priors on the weights of RNNs fit to neural data, and optimize the initial state to maximize the trajectory separation between the different RNNs. This initial state is our proposed optimal perturbation. We demonstrate the practicality of our framework by applying it to distinguish different proposed models of recurrent dynamics underlying short-term memory. The last part of dissertation is concerned with whether and how recurrent computation can facilitate recognition of objects under occlusion, which has been considered a challenging computational task. An intuitive suggestion for solving this task has been the explain-away computation, whereby the altered appearance of the occluded object is explained away as due to the occluder. We show that this explain-away computation arises naturally in recurrent, but not feedforward, models of the visual pathway, without any constraints placed on them. This suggests that the explain-away computation may be a type of neural computation that recurrent connections can uniquely perform. In summary, these three studies shed light on what kind of computations recurrent connections can perform, how they perform these computations, and how we can distinguish distinct recurrent models of a neural circuit through perturbations.
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 | 2022; ©2022 |
| Publication date | 2022; 2022 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Kang, Byungwoo |
|---|---|
| Degree supervisor | Druckmann, Shaul |
| Degree supervisor | Ganguli, Surya, 1977- |
| Thesis advisor | Druckmann, Shaul |
| Thesis advisor | Ganguli, Surya, 1977- |
| Thesis advisor | Petrosian, Vahe |
| Degree committee member | Petrosian, Vahe |
| Associated with | Stanford University, Department of Physics |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Byungwoo Kang. |
|---|---|
| Note | Submitted to the Department of Physics. |
| Thesis | Thesis Ph.D. Stanford University 2022. |
| Location | https://purl.stanford.edu/yr302dc3601 |
Access conditions
- Copyright
- © 2022 by Byungwoo Kang
- 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...