Abstract
Understanding and constructing brain communications that capture dynamiccommunications across multiple regions is fundamental to modern systemneuroscience, yet current methods struggle to find time-varying region-levelcommunications or scale to large neural datasets with long recording durations.We present a novel framework using Markovian Gaussian Processes to learn braincommunications with time-varying temporal delays from multi-region neuralrecordings, named Adaptive Delay Model (ADM). Our method combines GaussianProcesses with State Space Models and employs parallel scan inferencealgorithms, enabling efficient scaling to large datasets while identifyingconcurrent communication patterns that evolve over time. This time-varyingapproach captures how brain region interactions shift dynamically duringcognitive processes. Validated on synthetic and multi-region neural recordingsdatasets, our approach discovers both the directionality and temporal dynamicsof neural communication. This work advances our understanding of distributedneural computation and provides a scalable tool for analyzing dynamic brainnetworks.