Decoding Large Scale High Density Neural Recordings in Non-Human-Primates using a Spiking Neural Network

Abstract

To understand the complex processes underlying perception and action, it is necessary to understand the encoding mechanisms of information in the brain. Decoding, the mapping of brain responses back to a stimulus, is a useful tool for investigating what information is present in brain signals and in what format. Traditional approaches rely on firing-rate-based machine-learning models. However, previous research suggests that spike timing provides additional information. This study investigated the contribution of spike timing to the encoding of information in neuronal spike trains. This is achieved by decoding high-density large-scale neural spike trains with high spatiotemporal resolution recorded with a Neuropixels probe. The type of recording used eliminates the need for population or trial averaging, thus leaving temporal information intact. To assess the presence of informational spike time encoding, the decoding performances of a spiking neural network (SNN) and artificial neural network (ANN) using spike trains directly were assessed. As a baseline comparison, a support vector machine (SVM) model using average firing rates as input was included as well. The impact of inspecting the temporal dynamics of spike trains on orientation decoding performance and what this reveals about the contribution of spike timing to neural information transfer was investigated. The study revealed that incorporating spike timing into neural decoding improves performance, supporting the temporal and dual coding paradigms for information processing. However, further studies are needed to validate this.