This repository provides the implementation of SBIND (Spatiotemporal Behavior modeling in Imaging Neural Data), a deep learning framework for modeling raw neural imaging data.
Mohammad Hosseini and Maryam M. Shanechi. Dynamical Modeling of Behaviorally Relevant Spatiotemporal Patterns in Neural Imaging Data. In Proceedings of the 42nd International Conference on Machine Learning (ICML), 2025.
The following notebook contains usage example for SBIND:
The following are the key classes used to implement the SBIND model based on the formulation explained in the paper.
-
CONVSBIND(./sbind/convsbind.py): This is the main SBIND model class. It integrates the two ConvRNN modules (ConvRNN1 for behaviorally relevant dynamics and ConvRNN2 for other neural dynamics) and implements the full two-phase learning process described in the paper. -
SBINDTrainer(./sbind/sbind_trainer.py): This class is a utility trainer that contains the functions to fit the SBIND model, generate predictions on new data, and run validation. It handles the training loops, optimization, and saving/loading of the model.
Copyright (c) 2025 University of Southern California
See full notice in LICENSE.md
Mohammad Hosseini and Maryam M. Shanechi
Shanechi Lab, University of Southern California