Neural computational mechanism of compositional generalization

Thesis

This thesis hopes to provide further understanding of the representation subserving compositional generalization: how reusable knowledge is acquired from experience, how it is represented in the brain, and what are the computations over the representations that support generalization in novel situations. In Chapter 1, I presented an overview of the literature and the theoretical framework adopted by this thesis. Chapter 2 introduced the task paradigm used by this thesis. This task required participants to learn how to predict locations of treasure based on coloured animals from a subset of training stimuli with feedback and generalize the learned rules to a held-out subset of stimuli without feedback. In an fMRI study, I identified the variations in the dimensionality and orthogonality of the rule representations across hippocampus, ventral medial prefrontal cortex, posterior parietal cortex and primary visual cortex that support successful compositional rule generalization. Building on these variations, Chapter 3 reported simulation studies to explore the cost and benefit of compressing the high-dimensional representations of discrete features to solve the task. Given the design limitations of the fMRI study in Chapter 2, I conducted a series of pilot experiments in Chapter 4 that gradually improved the design for extending the task to examine cross-context generalization. In this extension, I showed that participants could combine the rules learned in different temporal contexts to make novel inferences. After characterizing the properties of abstract representations and how they are put to use in a compositional fashion to solve new problems, the thesis turns to the learning process underlying the formation of abstract representation. I further extended the task design to investigate structural transfer and how some curricula facilitated faster learning and better transfer performance than others in Chapter 5. In this task, participants were taught stimuli-label associations which were governed by a hidden relational structure. Using computational simulations, I showed that abstract representation of the hidden structure is crucial for successful generalization. In a behavioural study, I showed that: 1) participants could utilize the hidden relational structure to make predictions when probed on a new set of labels; 2) training curricula that made features that were relevant for the relational structure vary slowly in time could promote the factorised representation of structure and facilitate transfer.

Authors: Liang, Z

• DOI: 10.5287/ora-dm1k6kyj5
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: fMRI; compositionality; generalization; neural geometry; representation learning
• Subjects: Cognitive neuroscience