Computational Investigations of Human Creativity

Surabhi S Nath, Peter Dayan

Creativity is an important, yet elusive, human characteristic. Despite strides in understanding creativity as a cognitive ability, there is a paucity of computational studies associated with it. Our work aims to fill this gap by investigating the mechanisms underlying human little-c creative artistic work, from a reward learning perspective.

Extending work by Hart et al. (2017) and Rafner et al. (2023) studying creativity production using the creative foraging game, we fashion a tractably constrained experimental setting involving 5 x 5 binary pixel patterns. Patterns in this space belong to various classes (example classes shown in Fig a.1) and are connected by moves (Fig a.2).

We define a relevant taxonomy (Fig b) distinguishing two types of creativity--static (creativity of a pattern) and dynamic (creativity of a pattern creativity of a move); and two modes of creativity--evaluation and production. We outline various possible underlying computational mechanisms such as (1) an immediate value function based on utility and novelty of the pattern driving static creativity, (2) a long-run value function based on future affordances of a move driving dynamic creativity, (3) the history-dependent nature of evaluation and (4) a search process guiding the production.

We develop a series of behavioural experiments including creativity rating tasks to rate individual or sequences of patterns, and creativity production tasks to produce patterns with or without constraints.

Broadly, through this work, we aim to enrich the computational understanding of creativity.