Grid-like codes emerge in an LSTM path-integrator and are causally load-bearing

Banino et al. train an LSTM path-integration network on foraging trajectories, then feed its representation to an A3C deep RL navigation agent [banino-etal-2018-vector-based-navigation] Gridness-score analysis of spatial autocorrelograms confirms hexagonal grid-cell-like and border-vector-cell-like periodic tuning emerges purely from training [banino-etal-2018-vector-based-navigation] The resulting agent exhibits vector-based navigation and shortcut-taking behavior [banino-etal-2018-vector-based-navigation] Ablating the highest-gridness units significantly degrades navigation while ablating non-grid units does not (effect sizes corroborated via code documentation, not the paywalled Nature text) [banino-etal-2018-vector-based-navigation]