Data-driven system identification and neural circuit modelling of perceptual decisions

Abstract

Perceptual decision-making involves transforming sensory information into a singular choice. Although numerous cognitive computational models have been developed to explain experimental data, robust data-driven identification of such models from neural activity remains limited. Additionally, classical perceptual decision tasks often confound sensory processing with motor actions, leaving the neural mechanisms underlying flexible, motor-independent perceptual decisions poorly understood. Recent use of high-resolution neuronal recordings in novel memory-based sampling paradigms have begun to reveal related neuronal correlates, yet these neurons exhibited puzzling, delayed decision-related activity, and the possible underlying neurocomputational mechanisms remain unclear.

This thesis addresses these challenges through three key contributions:

First, an adapted data-driven symbolic regression approach, Sparse Identification of Nonlinear Dynamics (SINDy), is employed to identify stochastic decision models by directly uncovering the underlying dynamical equations. This demonstrates that latent decision dynamics can be reliably recovered from noisy and partially observable first-passage time data.

Second, through advanced signal processing and statistical re-analysis of non-human primate neuronal data from tasks that dissociate perceptual decisions from overt motor plans, the thesis enhances identification of neuronal correlates linked to decision memory-based actions, elucidating the diverse neuronal responses involved in flexible perceptual decision-making.

Third, a biologically based mean-field neural circuit model is developed to rigorously explain the observed re-analysed neuronal dynamics and decision behaviours from the re-analysis. Moreover, this model generalises across different decision task contexts, including those involving cognitive interference, demonstrating that a canonical neural circuit suffices to yield mechanistic insights into flexible perceptual decisions without the necessity of (re-)learning.

Together, these studies propose a unifying computational perspective for decision neuroscience, spanning data-driven symbolic machine learning for uncovering underlying decision model, neural data science for identifying decision-related neurons, and theoretical modelling of neural circuit mechanisms underlying flexible decision behaviour.

Date of Award	Jun 2026
Original language	English
Sponsors	Department for the Economy
Supervisor	Kongfatt Wong-Lin (Supervisor) & Saugat Bhattacharyya (Supervisor)