TRACE: Contrastive learning for multi-trial time series data in neuroscience
Published in Advances in Neural Information Processing Systems (NeurIPS 2025) 38, 2026
This work presents TRACE, a self-supervised contrastive learning method for multi-trial time series data in neuroscience, which averages across different subsets of trials to generate positive pairs and creates two-dimensional embeddings. The approach demonstrates superior performance on simulated data and captures biologically relevant continuous variation and cell-type-related cluster structure in neural recordings.
Recommended citation: Schmors, L., Gonschorek, D., Böhm, J. N., Qiu, Y., Zhou, N., Kobak, D., Tolias, A., Sinz, F., Reimer, J., Franke, K., Damrich, S., & Berens, P. (2026). "TRACE: Contrastive learning for multi-trial time series data in neuroscience." Advances in Neural Information Processing Systems (NeurIPS 2025). 38, 79837-79866.
Download Paper