Abstract
Neuronal systems need to process temporal signals. We here show howhigher-order temporal (co-)fluctuations can be employed to represent andprocess information. Concretely, we demonstrate that a simple biologicallyinspired feedforward neuronal model is able to extract information from up tothe third order cumulant to perform time series classification. This modelrelies on a weighted linear summation of synaptic inputs followed by anonlinear gain function. Training both - the synaptic weights and the nonlineargain function - exposes how the non-linearity allows for the transfer of higherorder correlations to the mean, which in turn enables the synergistic use ofinformation encoded in multiple cumulants to maximize the classificationaccuracy. The approach is demonstrated both on a synthetic and on real worlddatasets of multivariate time series. Moreover, we show that the biologicallyinspired architecture makes better use of the number of trainable parameters ascompared to a classical machine-learning scheme. Our findings emphasize thebenefit of biological neuronal architectures, paired with dedicated learningalgorithms, for the processing of information embedded in higher-orderstatistical cumulants of temporal (co-)fluctuations.