Human understanding: behaviour, brain and neural computation

What does it mean for an agent to understand the world? This problem is central to diverse fields, from philosophy to AI research. Here, we study human understanding from the standpoint of behaviour and neural computation. We combine deep learning theory, large-scale behavioural testing, and neuroimaging techniques that examine how behaviour and neural coding adapt as learners transition from naivety to understanding on complex tasks. We start from the premise that understanding requires knowledge to be appropriately structured and composed into new mental models. We focus on the ways that neural representations change as new knowledge structures (e.g. semantic hierarchies) are acquired. We will study the discrete stages (and moments of insight) through which learning passes as humans gradually master a task. We will ask why temporal structuring of information (curriculum learning) benefits human learning, and explore changes in representational format in human brain signals that occur as humans compose new knowledge from existing building blocks. In each of these projects, we test predictions motivated by simulations involving deep neural networks. Finally, we will collect large- scale data from people all over the world solving concept learning tasks, and model their learning trajectory with modern machine learning tools.