| Summary: | <p>Humans have a remarkable capacity to understand and act in situations they have not encountered before. This capacity relies on clever representations of what is common across experience. Relational representations are particularly powerful for this purpose, because they allow for composing novel situations from familiar building blocks. In this thesis, I use neuroimaging and computational modelling to investigate how the brain constructs knowledge and plans.</p>
<p>In a functional magnetic resonance imaging experiment I study the origin of knowledge about objects and situations. Bar stools look like tables, but are conceptually more similar to sofas. What is the source of these abstract similarities? I hypothesise that the compositional regularities of experience induce meaning in objects and situations, reflected in brain activity along the temporal to the frontal lobe. To test this hypothesis, I designed a paradigm that probes neural representations from shape to understanding. In a controlled experiment with precise quantitative neural predictions, I do not find convincing evidence of the hypothesised abstract representations in the brain.</p>
<p>Through a computational model I propose a mechanism for constructing plans. When navigating a new city, we know that to visit a sight on the other side of the river, we have to find a bridge first. I show how building cognitive maps from relative positions of such common structural elements allows to immediately reuse previously learned plans. Object vector cells in entorhinal cortex are an attractive neural substrate to represent these relative positions, because they imply actions and path integrate. I propose a central role for replay in constructing the compositional map in hippocampus. The resulting model generalises plans, suppresses path integration noise, and extends naturally to hierarchical tasks.</p>
<p>This thesis thus suggests a central role for relational representations and composition in making sense of the world and making decisions in it.</p>
|
|---|