Item

Abstract

Biological agents are adept at flexibly solving a wide range of cognitively challenging decision-making problems given woefully
little experience. This capacity rests on one fact about the problems themselves: that there is substantial recurring
structure; and two facts about us: that we can extract the structure and build internal representations of it based on the
statistics of observations, and that we can use those representations when solving new tasks. Artificial agents could benefit
from copying these characteristics.
An important form of statistical structure is a hierarchy. We therefore investigated the formation of hierarchical representations
in human subjects using a novel, sophisticated, shape composition task, in which subjects learn how composite shapes
are formed from a restricted set of basic building blocks. Understanding a new shape in these terms has been shown to involve
a form of internal, imagined, construction process. The task involved hierarchical structure with certain pairs of building
blocks tending to co-occur as hierarchical ’chunks’. Picking up on these chunks would facilitate the task of understanding
new shapes. We found that subjects learnt and employed hierarchically structured representations when composing visual shapes. Further,
we found that subjects generalised these structured representations to unseen stimuli. Subjects correctly identified previously
unseen shapes that contained hierarchical structure to be more likely to be part of the training set compared to random shapes
with no hierarchical structure. Further, when asked to complete novel shapes, subjects relied on hierarchical structure to
generate solutions.
Taken together, this suggests humans possess strong inductive biases for learning, employing, and generalising hierarchical
structures in visual understanding. The computational and neural bases of these capacities are not yet clear.