INTRODUCTION: Cognitive tasks with varying levels of complexity can be solved efficiently by biological neural networks. However, we still lack understanding of how structural and functional features of these networks are affected by increasing the complexity of the goal function.

AIM(S): Here we address this open problem by investigating the emergence of minimal, overrepresented neural network motifs in a classic Atari game. Using a Genetic Algorithm to evolve ensembles of Artificial Neural Networks (ANN) to successfully play Pong we found the simplest network motif to consist of excitatory and inhibitory edge integrating the ball and paddle position into a single output node. Surprisingly, we found that this policy is successful for varying degrees of velocities and different paddle sizes.

METHOD(S): To answer whether more complex motifs occur as cognitive demands increase, we evolved additional ANNs under a version of Pong that introduces uncertainty by uniformly sampling angular velocity of the ball on every wall collision. Two ensembles were evolved, acting as Low- and High-Noise variants, and a dimensionality reduction of the structural features of successful ANNs was performed, followed by clustering and subgraph occurrence analysis. The successful ANNs were then re-evaluated under the same task using Wilson Lower Confidence Bound test to further filter optimal policies.

RESULTS: We made three observations: 1) the increased uncertainty produced consistently more complex motifs than earlier ensembles without uncertainty; 2) the minimal motif occurs as a subgraph in all other solutions; however, 3) in every ensemble one cluster consistently featured approximately 40% of non-standard solutions, that is, solutions that did not contain the simplest motif.

CONCLUSIONS: Hence, we speculate that increasing cognitive load results in the emergence of new network motifs that are qualitatively different from simple functional solutions.

FINANCIAL SUPPORT: This research was supported by the National Science Center, Poland (SONATA BIS, 2021/42/E/NZ2/00188)