Back to search

Boston University

Noise and inductive bias in generalization: insights from humans, rodents, and computational models of decision-making and abstract reasoning

Abstract

dc:description.abstract

This dissertation investigates how variability (“noise”) and priors (“inductive biases”) interact to support—or impair—generalization across biological and artificial systems. Chapter 1 (Introduction) situates this inquiry within existing theory and empirical work in cognitive science, neuroscience, and machine learning. It reviews classic and contemporary models of decision making and abstract reasoning, defines key constructs (noise, inductive bias, generalization), and outlines the overarching research questions and methodological approach that guide the subsequent empirical and computational studies. Chapter 2 develops a novel online, nonverbal evidence-accumulation game for humans inspired by pulse-based rodent tasks. Across hundreds of participants trained either with explicit instructions or solely via reward-based feedback, both groups learned equally rapidly and adopted comparable decision strategies. Behavioral modeling revealed that internal noise scales nonlinearly with evidence—mirroring animal findings but contradicting classical diffusion accounts of human perception. These results challenge prevailing theories of human evidence accumulation and demonstrate the value of feedback-driven paradigms for studying populations unable to rely on verbal instructions. Chapter 3 adapts the evidence-accumulation game for adolescents with autism spectrum disorder (ASD) and their typically developing siblings. ASD participants exhibited slower learning and altered perceptual integration, correlating with social and adaptive symptom severity. Critically, artificial neural networks endowed with increased noise reproduced these ASD-like deficits, suggesting a mechanistic link between neural variability and learning impairment. Chapter 4 establishes a rat experimental paradigm of abstract generalization using the same pulse-based evidence-accumulation task. Rats trained under curricula that restrict stimulus–action pairings generalized to novel pulse sequences according to a hybrid Exemplar-to-Prototype learning model. In this framework, internal noise broadens perceptual experience, facilitating transfer beyond the training set. Ongoing pilots with EEG and optogenetic perturbations aim to uncover the neural circuits mediating these effects. Chapter 5 presents an algorithmic one-shot rule-learning model for Raven’s Progressive Matrices–style tasks. Applied to a simplified Raven’s task designed for human behavioral and neuroimaging studies, the model achieves near–human-level performance in the symbolic condition, expresses discovered rules explicitly, and generalizes to perceptual problems with continuous patterns. These results highlight how algorithmic inductive biases can yield rapid abstraction and inform both cognitive theory and the design of more flexible AI systems. Chapter 6 introduces a Graph Neural Network (GNN) framework to capture the abstract representations underlying human reasoning. Tested on a modified Abstraction and Reasoning Corpus (ARC) with human subjects, the GNN ingests graphs encoding multiple relationship types and levels of abstraction, successfully reproducing human solution patterns. By exploring the learned representation space, we identify topological features that predict both reasoning success and characteristic error modes, providing a bridge between representational geometry and circuit-like implementations. Conclusions: Together, these six chapters illuminate how noise and inductive bias jointly promote—or impede—flexible generalization across humans, rodents, and machines. By integrating cross-species behavioral paradigms, neurophysiological measurements, algorithmic and statistical models, and deep neural networks, this work advances our understanding of the mechanisms that enable robust decision-making and abstract reasoning in both typical and neurodivergent brains and guides the development of more adaptable artificial agents.

Degree

thesis:*
Name thesis:degree_name
Doctor of Philosophy
Level thesis:degree_level
doctoral
Discipline thesis:degree_discipline
Neuroscience
Grantor
Boston University
Year dc:date.issued
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Do, Quan
Advisors dc:contributor.advisor
  • Hasselmo, Michael E.
  • Scott, Benjamin B.
ORCID dc:identifier.orcid

Subjects

dc:subject × 1

Rights

dc:rights
Statement dc:rights
  • Attribution-NonCommercial-NoDerivatives 4.0 International
Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/2144/53179

Chain of custody

source
Harvested from
Boston University
Base URL
open.bu.edu/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
related terms
citation

Do, Quan. Noise and inductive bias in generalization: insights from humans, rodents, and computational models of decision-making and abstract reasoning. doctoral thesis, Boston University, 2025. https://hdl.handle.net/2144/53179