Representational and Temporal Dynamics in Neural Decoding: Linear and Nonlinear Models for Position and Velocity Prediction

Minsung Cho, Jaesung Yoo, Stefan M Lemke, Jian Zhong Guo, Adam W Hantman

Keywords: Behavior Decoding, Deep Learning, Computational Neuroscience

Abstract: Understanding how neural activity encodes behavior remains a central challenge in systems neuroscience. Neural spike trains offer high temporal resolution and may contain information about specific motor features, but the extent of this information remains unclear. In this study, we evaluate how well motor features can be predicted from spike trains using two type of decoders: linear regressison and a deep neural network. Using data from mice performing a reach-to-grab task, we compare predictions of hand position and two velocity representations. To assess whether models capture meaningful patterns rather than superficial correlations, we introduce artificial temporal lags between neural and behavioral data. This disrupts genuine associations and reveals whether decoding reflects true information content. Decoding score across lags show how behavioral information is distributed in spike trains across time and how linear and nonlinear models decode such information. This approach provides a principled framework for evaluating the behavioral relevance of neural activity.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 12011