Learning efficiently in uncertain and structured worlds

Thesis

Intelligence emerges from the interaction between learning mechanisms and the environments that shape them. This thesis investigates how biological and artificial systems navigate two fundamental challenges: learning under uncertainty when information is scarce or costly, and exploiting structured environments through abstraction and generalization.

The thesis develops both methodological innovations and empirical insights across three interconnected investigations. First, I introduce a hybrid modeling framework that combines theory-driven cognitive models with data-driven artificial neural networks, using symbolic regression to transform learned neural network functions into interpretable mathematical expressions. This approach bridges the flexibility of machine learning with the interpretability required for scientific understanding, providing a general method for discovering computational principles underlying cognition.

Second, I apply this framework to understand how humans compute the value of information during decision-making under uncertainty. Using ultra-high field 7T fMRI, I identify distinct neural signatures across neuromodulatory nuclei and cortical regions. The ventral tegmental area balances exploration versus exploitation by encoding opposing signals for information value and selection value. The anterior insula and anterior cingulate cortex guide the information sampling strategy. Symbolic regression reveals that the value of information follows exponential functions integrating evidence from both attended and unattended options, with parameters that capture individual differences in exploration strategies. These equations generalize to predict behavior in independent explorationexploitation tasks.

Third, I investigate computational mechanisms underlying rapid generalization across sequential learning tasks. I develop computational models comparing baseline recurrent neural networks against architectures augmented with grid-hippocampal episodic memory systems. Models implementing abstract 2-dimensional maps through grid cell path integration achieve dramatic learning acceleration comparable to animal behavior. Critically, backward temporal credit assignment through causal attribution and episodic binding enables near-instantaneous transfer after initial learning. These computational findings generate specific neural predictions testable through fMRI pattern analysis and transcranial ultrasound stimulation.

Together, these investigations provide evidence consistent with the idea that intelligence may emerge from structured representations that facilitate rapid learning, effective information sampling, and flexible generalization across contexts.

Authors: D'Ambrogio, S

• DOI: 10.5287/ora-xmeardm88
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: reinforcement learning; exploration–exploitation; abstraction; decision-making; grid cells; generalization
• Subjects: Cognitive neuroscience; Computational neuroscience