Goal-Directed Navigation in Multi-Scale Environment Based On Wavefronts Propagation in Cortical Network

Goal-directed navigation is a field receiving much research attention. One of the research branches involves bottom-up construction of navigation ability based on artificial neural network inspired by neurophysiology. Examples include prefrontal cortical network, which organizes information from place cells, motor cells and other interneurons, providing reward calculation and direct action readout in a single framework. Although this kind of models works fine in small-scale environment, it fails in large-scale environment. In order to solve this problem, we introduced wavefronts propagation method in the cortical network, which is a multi-scale planning method with low time complexity. The main parts of our model are reward cells generating non-decaying wavefronts, and interneurons recruiting Spike Timing-Dependent Plasticity (STDP) for storing preferred direction. Tolman maze tests showed that robot recruiting this model has spatial cognition similar to that of rats in Tolman's experiments, providing a possible cognitive map formation mechanism Tolman proposed. Moreover, our model is both scale-free and robust to neuron noise.