A study on interpreting spiking predictions

This study presents an information theoretic evaluation of interpretation (or decoding) methods commnly used for making predictions with spiking neural networks. It is not currently known how spiking neural network outputs (i.e. spike trains) encode information. Thus, when making predictions with spiking neural networks several decisions have to be made: how information is being encoded (e.g. number of spikes), how that information can be interpreted into a prediction (e.g. interpretation method), and how to represent that prediction (e.g. firing rates). In most cases these decisions are entirely subjective. While many thechniques have been proposed for measuring the amount of information contained in output spike trains about the problem characteristics (i.e. class), such as study has not been conducted for evaluating and comparing interpretation methods. To address this, we conduct a study that evaluates popular interpretation methods in terms of the amount of (Shannon) information they capture from the spike trains being interpreted. To enable this comparison we require a consistent representation for predictions. Thus, we demonstrate how several popular interpretation methods can be modified to make predictions using the same representation. Based on the measured information loss, we draw conclusions regarding the benefits and drawbacks of different interpretation methods.