Abstract
Drawing parallels between Deep Artificial Neural Networks (DNNs) andbiological systems can aid in understanding complex biological mechanisms thatare difficult to disentangle. Temporal processing, an extensively researchedtopic, is one such example that lacks a coherent understanding of itsunderlying mechanisms. In this study, we investigate temporal processing in aDeep Reinforcement Learning (DRL) agent performing an interval timing task andexplore potential biological counterparts to its emergent behavior. The agentwas successfully trained to perform a duration production task, which involvedmarking successive occurrences of a target interval while viewing a videosequence. Analysis of the agent's internal states revealed oscillatory neuralactivations, a ubiquitous pattern in biological systems. Interestingly, theagent's actions were predominantly influenced by neurons exhibiting theseoscillations with high amplitudes and frequencies corresponding to the targetinterval. Parallels are drawn between the agent's time-keeping strategy and theStriatal Beat Frequency (SBF) model, a biologically plausible model of intervaltiming. Furthermore, the agent maintained its oscillatory representations andtask performance when tested on different video sequences (including a blankvideo). Thus, once learned, the agent internalized its time-keeping mechanismand showed minimal reliance on its environment to perform the timing task. Ahypothesis about the resemblance between this emergent behavior and certainaspects of the evolution of biological processes like circadian rhythms, hasbeen discussed. This study aims to contribute to recent research efforts ofutilizing DNNs to understand biological systems, with a particular emphasis ontemporal processing.