Thorugh the Recurrent Neural Network Looking Glass: Structure-Function Relationships in Cortical Circuits for Predictive Coding
Author(s)
Balwani, Aishwarya
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
The elucidation of structure-function relationships in neural circuits forms the cornerstone of theoretical and computational neuroscience. However, despite decades of neuroscientific inquiry and technological advancements, our grasp on how neural architecture shapes its computations remains incomplete. This limitation is particularly evident in our understanding of the cortex, where its intricate laminar structure is arranged into the recurring neuroanatomical motif known as the canonical cortical microcircuit. While we now have extensive knowledge of the structural scaffolding itself, including its cell types and connectivity patterns, we lack a comprehensive understanding of how this architecture supports its sophisticated information processing. Specifically, the precise mechanisms by which the canonical microcircuit and its inter-areal connections facilitate predictive coding, a theoretical framework positing that the cortex implements hierarchical, predictive computations for efficient perception and inference, remains elusive.
While in-vivo efforts to experimentally test the underlying mechanisms of predictive coding still face enormous challenges, in-silico approaches present an encouraging avenue to study the phenomenon. Specifically, recurrent neural networks (RNNs) offer a powerful framework to investigate hypotheses about hierarchical predictive coding, owing to their ability to effectively exploit temporal dependencies typical of natural stimuli, learn distributed representations, and model complex, multi-regional neuronal dynamics. This thesis subsequently employs RNNs as models of cortical circuits, illustrating how they can be leveraged to study the inherent architectural biases of the canonical cortical microcircuit and their functional implications, in light of the predictive coding hypothesis. Furthermore, this dissertation also develops tools that address limitations concerning the anatomical fidelity of conventional RNN architecture and training, hence improving their overall biological plausibility and potential for modelling real neural circuits to gain neuroscientific insights. The contributions of this thesis therefore provide a robust, computational framework for modelling and interpreting complex, multi-regional neuronal data, ultimately paving the way for deeper insights into how the structural properties of cortical circuits support sophisticated information processing and intelligent behavior.
Sponsor
Date
2025-04-23
Extent
Resource Type
Text
Resource Subtype
Dissertation