Neural sampling from cognitive maps supports goal-directed planning and imagination

Imagining plans for solving problems is a cornerstone of our higher cognition. We can even create plans for reaching goals that we never encountered before. However, the ways the brain addresses goal-directed imagination is largely unknown and current AI methods provide limited solutions to this problem. Here, we introduce a novel computational model – the generative cognitive map learner (GCML) – that successful addresses goal-directed imagination and provides a new hypothesis about the brain mechanisms supporting it. The GCML uses stochastic samples from learned cognitive maps – data structures in which brains encode learned relations between actions and states – to form trajectories towards both known and novel goals. In a series of simulations, we show that the GCML provides surprisingly effective heuristic solutions to spatial navigation tasks, generic problem solving tasks, and compositional tasks that require generating imagined trajectories to novel goals. The sampled trajectories capture key characteristics of hippocampal replay activity, which provide a candidate mechanism for goal-directed imagination in the brain. Since the GCML only requires simple biologically plausible local synaptic plasticity and shallow neural networks, it can be readily implemented in energy-efficient neuromorphic hardware.