Understanding biological active sensing behaviors by interpreting learned artificial agent policies

Sonja Johnson-Yu, Satpreet Harcharan Singh, Federico Pedraja, Denis Turcu, Pratyusha Sharma, Naomi Saphra, Nathaniel Sawtell, Kanaka Rajan

Keywords: artificial agents; recurrent neural networks; active sensing; weakly electric fish; deep reinforcement learning

TL;DR: Our electric fish agent shows emergent behavior modules that provide intuitions about latent variables governing behavior.

Abstract: Weakly electric fish, such as Gnathonemus petersii, generate pulsatile electric organ discharges (EODs) that enable them to sense their environment through active electrolocation. This plays a crucial role in several key behaviors, such as navigation, foraging, and avoiding predators. While the anatomical and physiological organization of the active electrosensory system has been extensively studied, the contribution of active electrolocation to adaptive behavior in naturalistic settings remains relatively underexplored. Here we present a preliminary in silico model of active sensing in electric fish, using a neural network-based artificial agent trained by deep reinforcement learning to perform an analogous active sensing task in a 2D environment. The trained agent recapitulates key features of natural EOD statistics, shows emergent behavioral modularity, and provides intuitions about the representation of key latent variables governing agent behavior, such as energy levels (satiety).

Submission Number: 3