Flexible Dimensionality of Representational Spaces in Category Learning

Introduction

Our visual system frequently has to classify complex, high-dimensional inputs. A key learning objective of the brain is thus to identify diagnostic dimensions. Often, tasks require simultaneous consideration of multiple dimensions. Yet, learning many dimensions is computationally challenging.

Research Question

Here, I ask how the visual system tackles the challenge of learning high-dimensional tasks. Some theories suggest that the brain does so by compressing dimensions, while others suggest dimensionality expansion. Yet, dimensionality compression and expansion both have advantages and disadvantages, and some studies find dimensionality compression where others find expansion.

Hypothesis

This raises the hitherto unanswered question of what determines whether the brain invokes either of the two strategies. I hypothesize that instead of settling on a single strategy, the brain can reap the benefits of dimensionality compression and expansion by flexibly adjusting dimensionality to the task at hand. This entails the novel prediction of flexible neural codes that can switch dimensionality.

Methodology

To test this theory, I build on a multimodal, multispecies approach I have developed to study learning. The following steps will be taken:

1. Using the paradigmatic case of visual category learning, I will establish the effect of task dimensionality on the structure of mental representations in behavior.
2. I will determine how task dimensionality transforms neural activity using neuroimaging in humans.
3. I will identify the neural building blocks of flexible dimensionality using electrophysiology and causal perturbations in rhesus monkeys.
4. I will unravel computational principles of flexible dimensionality with artificial neural networks.

Significance

This combination of species and techniques is ideally suited to unravel the neural mechanisms for coping with high-dimensional tasks. By elucidating the flexibility of mental and neural representations, I aim to reveal a hitherto unknown principle governing learning and stimulate future educational applications.