A Bergson-Inspired Adaptive Time Constant for the Multiple Timescales Recurrent Neural Network Model

Thomas F Burns, Fabien CY Benureau, Jun Tani

Abstract: We introduce an adaptive time constant for the multiple timescales recurrent neural network (MTRNN) inspired by the work of philosopher Henri Bergson (1859-1941). Observed variations in the recent activity of MTRNN neurons is used to adapt the time constant. We analyse how the time constant adapts in response to neuronal activity for a simple one-dimensional function-fitting task.