Path classification by stochastic linear recurrent neural networks

Abstract We investigate the functioning of a classifying biological neural network from perspective statistical learning theory, modelled, in simplified setting, as continuous-time stochastic recurrent (RNN) with identity activation function. In purely (robust) regime, we give generalisation error bound that holds high probability, thus showing empirical risk minimiser is best-in-class hypothesis. show RNNs retain partial signature paths they are fed unique information exploited for training and classification tasks. argue these easy to train robust support observations numerical experiments on both synthetic real data. also trade-off phenomenon between accuracy robustness.