Conference on Cognitive Computational Neuroscience

Subicular neurons in freely moving animals exhibit spatial firing that is influenced by environmental geometry. Termed 'boundary vector cells' (BVCs), these neurons are posited to encode an animal's allocentric location relative to environmental boundaries. Because these spatial responses resemble those hypothesised as inputs to place cells, they are believed to shape CA1/3 activity. However, the subiculum is a major anatomical output of the hippocampus and thus a simple causal relationship is unlikely. Moreover, a subset of subicular neurons maintain trace firing fields after the relevant spatial feature has been removed, implying that these neurons contain a mnemonic component. Here we present an anatomically consistent, biologically plausible account of these subicular cells within the paradigm of the hippocampus as a predictive map. Specifically, we demonstrate that BVC-like activity patterns can arise from learning a successor representation using trajectories displaying realistic rodent biases towards boundaries. Our model uses hippocampal place cells as basis features to learn successor representations that recapitulate an animal’s proclivity to run along walls, thus forming elongated activity fields, statistically analogous to BVCs, that duplicate upon barrier insertion. Our model makes empirical predictions distinct from the traditional BVC framework, including subicular representations of corners and experimentally observed ‘boundary-off’ cells.