Unknown Input Reconstruction from Temporal Activity Patterns of Thermosensitive Neuronal Ensembles using Reservoir Computing*

Abstract

The paper addresses the problem of the unknown temperature reconstruction from the neuronal signal generated by networks of thermosensitive neurons modeled by the Hodgkin-Huxley formalism. We show that the instantaneous frequencies of the neurons do not provide sufficient information for the precise temperature reconstruction and that the temporal patterns of the instantaneous frequencies should be taken into account. For this purpose, we augment the considered network of thermosensitive neurons with a multi-layered artificial neural network (ANN) and train it to reconstruct the unknown piecewise-constant input temperature using a supervised learning technique. The input layer of the ANN receives the discretized (in time) trajectories of the instantaneous frequencies of neurons over a certain time interval. Interestingly more rich information containing the time evolution of membrane potentials in the input layer does not lead to any improvement of the unknown input reconstruction. This observation is in accordance with one of the fundamental principles of neuroscience stating that, except for a few highly specific contexts, information in neural systems is encoded in the temporal rather than voltage characteristics of action potentials. Finally, we benchmark the performance of the proposed reconstruction scheme against different types of input signals.