The best spike filter kernel is a neuron

Abstract

A common approach to extracting information from simulated spiking neural networks is to train readouts on a spike-rate variable obtained through convolution of output spike-trains with a filter. Here we argue that best practice is to use neurons as spike filters. We describe how neural circuits consist of stock and flow variables that co-determine each other and argue that membrane potentials provide access to the information contained in the circuit in a more natural and unbiased way than filtered spike-trains. We compare the two different approaches to readout calibration in a classification task.