Time cell encoding is decoupled from time perception in deep reinforcement learning agents
Ann Zixiang Huang, Dongyan Lin, Blake Richards, McGill University, Quebec AI Institute (Mila), Canada
Contributed Talks 1 Lecture
Grand Ballroom A-C
Fri, 26 Aug, 11:30 - 11:50 Pacific Time (UTC -8)
Time cells are shown to encode the unfolding of time by forming trial-consistent temporal receptive fields that are selective to particular moments. However, it remains unknown whether besides carrying temporal information, time cells directly contribute to the cognitive function of timing. Here, by training deep reinforcement learning (DRL) agents to compare the duration of two sequentially presented stimuli, we show that time cells naturally emerge and encode time elapsed regardless of the cognitive demand of timing. Furthermore, the temporal receptive field of individual cells does not rescale across different stimulus duration nor discriminates between correct and incorrect trials, suggesting a dissociation between time encoding and duration judgment in the DRL agent. Together, our findings posit that time encoding may emerge as an intrinsic circuit phenomenon of recurrent neural networks irrespective of the cognitive function of timing.