A T-maze LSTM forms place fields and switches manifold paths

Jude & Hennig train a 380-unit LSTM on a memory-dependent T-maze (sensory prediction plus Q-learning) with no explicit position or movement input [jude-hennig-2020-hippocampal-representations-maze] Place-field mapping (30%-of-peak threshold) finds well-isolated place fields covering the whole maze, a discrete point-attractor-like landscape [jude-hennig-2020-hippocampal-representations-maze] Substantial extrafield firing concentrates at cue and choice points in 56% of units after reward training, mirroring rodent CA1/CA3 recordings [jude-hennig-2020-hippocampal-representations-maze] UMAP shows the population trajectory abruptly switching from the left- to the right-trajectory manifold when the agent pauses at a choice point [jude-hennig-2020-hippocampal-representations-maze] No causal ablation of the geometric structure itself is performed [jude-hennig-2020-hippocampal-representations-maze]