Partial Information Decomposition for Time Series
Amanda Merkley
Abstract:
Of great interest in neuroscience is how information about a stimulus is distributed across multiple neurons and neural circuits. Partial information decomposition (PID) is an emerging topic in information theory that aims to decompose the information that multiple sources (i.e., neurons) carry about a target (i.e., stimulus). Particularly in the bivariate case, PID specifies the unique contribution of each source as well as what both convey redundantly and synergistically about the target. However, neural activity is almost always recorded as a dynamic process which creates an additional challenge for PID since most current measures are defined for random variables and do not account for temporal relationships. Through an example application on neural spiking data, we demonstrate how extensions of PID for time series can be applied to real data and illustrate the limitations of currently proposed measures.