View article

In this study, we propose a neural encoding method for fish motion learning using spiking neural network. The network is trained to associate a particular motion to its target response using our initially developed reward-based learning algorithm. For the encoding purposes, we use a recurrent neural network with sparse and random connection consisting of 1000 spiking neurons. Each point in a motion is represented by a group of neurons, in which a sequence of group stimulations forms a motion trajectory. The sequence is associated to a target response, represented by a group of response neurons. For this study, there are two groups of competing response neurons. In each learning trial, the sequence is activated and the network is rewarded or penalised depending on the winning response. The learning follows a simple and natural protocol implemented in a noisy and dynamic setting. Based on the experiment findings, the encoding method of fish motion trajectory seems feasible to be coupled with the reward based learning shown by the sequence recognition performance. Moreover, this is among the pioneer studies that implement motion trajectory learning using spiking neural network in a reward-based paradigm.