Fig. 2: Splitters emerge when dictated by task structure.

a Task environment for simulation (to mimic the experimental design of Zhao et al.¹⁸). The agent is presented with one of two possible cues, A or B (here represented as red vertical bars or black horizontal bars, respectively), before advancing through the track (C). It can then choose to go to either branch D or E to potentially receive a reward. b Fields are induced in model HPC cells via the induction of a “plateau potential” at a certain location after a certain cue. Half of the neurons are induced on trials which begin with cue A, and half are induced on trials which begin with cue B. c The cue-dependent paradigm, where location D contains reward if preceded by cue A, and location E contains reward if preceded by cue B. d The random reward paradigm, where locations D and E each result in a 50% chance of reward delivery, regardless of the cue shown. e Simulated population activity (i, iv), single-cell activity (ii, v), and experimentally recorded activity after learning (data from Zhao et al.¹⁸) (iii, vi), for both the cue-dependent task (i–iii) and the random reward task (iv–vi). Notably, the same induction protocol generates splitters in the cue-dependent case, and non-specific place fields in the random reward case. Red, trials which began with an A cue, black trials which began with a B cue. Shaded area represents SEM. n = 20 neurons for simulated activity and n = 12 cells for experimental data¹⁸.