An important issue in spatial memory is the learning of abstract representations suitable for navigation planning. To address this problem, we have already developed a planning system inspired by the columnar organization of the mammalian cortex [1]. This model provides a neuromimetic architecture capable of learning topological spatial representations and planning goal-directed actions. The work presented here deals with the ability to encode multiscale representations of the environment, in order to solve large maze tasks, This is shown by validating the model on a multiscale version of the Tolman & Honzik's detour task [2]. Simulation results demonstrate that the performances of the planning system are invariant with respect to the scale of the maze. A series of statistical analyses is provided to characterise the neural activities subserving spatial planning. It is shown that the structural properties of the environment are encoded by the discharges of the location-selective neurones of the model. Complementing this purely spatial coding, the activity of another class of neurones in the model integrates both spatial and reward-dependent information suitable for navigation planning.
展开▼
