Predictive Place-Cell Sequences for Goal-Finding Emerge from Goal Memory and the Cognitive Map: A Computational Model

Abstract

Hippocampal place-cell sequences observed during awake immobility often represent previous experience, suggesting a role in memory processes. However, recent reports of goals being overrepresented in sequential activity suggest a role in short-term planning, although a detailed understanding of the origins of hippocampal sequential activity and of its functional role is still lacking. In particular, it is unknown which mechanism could support efficient planning by generating place-cell sequences biased toward known goal locations, in an adaptive and constructive fashion. To address these questions, we propose a model of spatial learning and sequence generation as interdependent processes, integrating cortical contextual coding, synaptic plasticity and neuromodulatory mechanisms into a map-based approach. Following goal learning, sequential activity emerges from continuous attractor network dynamics biased by goal memory inputs. We apply Bayesian decoding on the resulting spike trains, allowing a direct comparison with experimental data. Simulations show that this model (1) explains the generation of never-experienced sequence trajectories in familiar environments, without requiring virtual self-motion signals, (2) accounts for the bias in place-cell sequences toward goal locations, (3) highlights their utility in flexible route planning, and (4) provides specific testable predictions.

A statistical note on the detection of autoassociative dynamics in hippocampal place-cell sequences

“…This experimental result contrasts with a recent theoretical model for the generation of replay sequences during SWRs, which showed a qualitative pattern of step-like activity transitions, associated with times of low population activity, although the underlying network structure was characerized by a graded pattern of synaptic weights, emerging from sequence learning (Jahnke et al, 2015). This discrepancy between experimental observations and theoretical results is further highlighted by our recent model (Gönner et al, 2017), which shows a qualitative pattern of step-like activity transitions despite continuous attractor dynamics, owing to the presence of strong population oscillations. Moreover, numerous theoretical models for the generation of place-cell sequences predict a spatially smooth structure of sequential activity whenever the spatial distribution of place fields is homogeneous (Redish and Touretzky, 1998, Molter et al, 2007, Hasselmo, 2009, Bush et al, 2010, Vladimirov et al, 2013, Chenkov et al, 2017.…”

“…By contrast, previous theoretical models predict that the spatiotemporal structure of place-cell sequences should reflect the distribution of place fields, typically observed to be spatially smooth. Motivated by this discrepancy between models and experimental data, we performed a quantitative comparison between these results and the spike trains generated by a network model for the generation of place-cell sequences (Gönner et al, 2017). Although the model is based on continuous attractor network dynamics, we observed that the movement of sequential place representations was phase-locked to the population oscillation, highly similar to the experimental data interpreted as evidence for discrete attractor dynamics.…”

Replay as wavefronts and theta sequences as bump oscillations in a grid cell attractor network

“…Other models have been proposed for theta phase precession [87][88][89][90] and replay [91][92][93][94][95][96]. For example, one grid cell model uses after-spike depolarization within a 1D continuous attractor network to generate phase precession and theta sequences [89].…”

“…Previously proposed replay models were all intended to address place cells, not grid cells. Several of them encode replay trajectories into synaptic weights, either through hard-wiring or a learning mechanism [94][95][96]. Two models have suggested that replays originate from wavefronts of activity propagating through networks of place cells [91,93].…”

Discovery of Hierarchical Representations for Efficient Planning

“…In order to perform (flat) BFS to find the shortest path from node s to node g, an external input can successively probe each neighbor of s by transiently activating the corresponding unit, triggering a "forward sweep" of activation that propagates through the circuit until it reaches and activates the unit of the goal state g. The neighbor of s that activates g in the least amount of time is the next node along the shortest path to g, so the agent can then physically transition to that state and repeat the process again and again, until finally reaching the goal state g. Assuming some form of short-term synaptic depression or ion channel inactivation that prevents the sweep from going backwards (Dobrunz et al, 1997), this implements precisely BFS. Similar schemes have been proposed to support forward trajectory planning in the hippocampus (Erdem and Hasselmo, 2012;Gönner et al, 2017).…”