Spatial representation of hierarchical recurrent neural network through predictive learning

Abstract

It is considered that the place cell is a fundamental neural basis for such spatial understanding. The previous studies proposed neural network models that can sustain the internal states corresponding to the spatial positions as attractors of dynamical systems. However, it is an open question how the place cell and spatial recognition can emerge from the subjective experiences of motion and sensory inputs without referring to the spatial position. To this question, we show that place-cell-like neurons are self-organized in our proposed hierarchical recurrent neural network which is trained to predict the future visuomotor states. It will be discussed about the spatial representation and neural network modelling in my talk.