8/ Our DRL agents also replicated the results from rat dorsolateral striatum by Toso and colleagues (doi.org/10.1016/j.ne...), which further suggested a dissociation between time encoding in these units and the behavior of time perception in rats and deep RL agents.

6/ Our results also hint that time cells might not simply be "time" cells: in freely-moving agents, they also encode space. They also encode sensory stimuli when there’s a demand to do so, suggesting that they encode not just *time* per se, but variables that unfold *over time*.

5/ However, if only the targeted cells are silenced while the activity of non-targeted cells is intact, the performance deteriorates slower than if the recurrent dynamics are corrupted, meaning time/ramping cells contribute to behavior as part of the recurrent dynamics.

4/ But, do these time-encoding units causally contribute to timing behavior? Yes and no. We show that lesioning time/ramping cells in-silico has the same impact to behavior as lesioning random units, meaning they don’t directly drive the downstream value and policy units.

3/ Why do these cells emerge? In timing tasks, they might be needed for time tracking. But in working memory tasks where there is no demand for time tracking, we show that learning to estimate the temporally discounted value function is what drove their emergence.

2/ How does the brain tell time? Neuroscientists have observed “time cells” and “ramping cells” whose activity correlates with elapsed time. We trained deep RL agents on simulated timing and time-dependent working memory tasks, and saw these cells naturally emerge in LSTM units.