Decoding Spike Trains from Neurons with Spatio-Temporal Receptive Fields

The point-process filter (PPF) is a real-time recursive algorithm that computes the minimum mean-squared error estimate of a behavioral state, given neural spiking observations. When used with stimulus-sensitive neurons that represent behavioral states transiently, the PPF needs to know the times at which stimuli will occur. However, these times will not be known a-priori. In this work, we develop a matched-filter point process filter (MF-PPF) that can decode behavioral states that are encoded transiently in neural activity when stimulus times are unknown. A linear filter matched to each neuron's temporal receptive field is used to estimate stimulus onset times, which are then fed into the PPF to decode the behavioral state. As an example, we use the MF-PPF to decode visual saliency from simulated superior colliculus spiking activity. This new decoder has the potential to decode behavioral states from brain regions with transient representations and temporal receptive fields.