Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
TRACE: Contrastive learning for multi-trial time series data in neuroscience
Authors: Lisa Schmors, Dominic Gonschorek, Jan Niklas Böhm, Yongrong Qiu, Na Zhou, Dmitry Kobak, Andreas S. Tolias, Fabian H. Sinz, Jacob Reimer, Katrin Franke, Sebastian Damrich, Philipp Berens
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that TRACE outperforms other methods, resolving fine response differences in simulated data. Further, using in vivo recordings, we show that the representations learned by TRACE capture both biologically relevant continuous variation, cell-type-related cluster structure, and can assist data quality control. |
| Researcher Affiliation | Academia | 1 Hertie Institute for AI in Brain Health, University of T ubingen, T ubingen, Germany 2 Institute for Ophthalmic Research, University of T ubingen, Germany 3 Centre for Integrative Neuroscience, University of T ubingen, Germany 4 Department of Neuroscience, Baylor College of Medicine, Houston, USA 5 Center for Neuroscience and Artificial Intelligence, Baylor College of Medicine, Houston, USA 6 Institute for Computer Science and Campus Institute Data Science, University of G ottingen, Germany 7 Department of Ophthalmology, Byers Eye Institute, Stanford University School of Medicine, USA 8 Stanford Bio-X, Stanford University, USA 9 Wu Tsai Neurosciences Institute, Stanford University, USA 10 Department of Electrical Engineering, Stanford University, USA |
| Pseudocode | No | The paper describes the methodology in detail within section 3 "TRACE: contrastive learning for multi-trial time-series data in neuroscience" but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures. |
| Open Source Code | Yes | Our code is available at https://github.com/berenslab/TRACE/tree/neurips25 and https://github.com/berenslab/TRACE_experiments. |
| Open Datasets | Yes | We used the Allen Institute Neuropixels visual coding dataset [36], which is part of the Allen Brain Observatory to test performance on a Neuropixels spiking dataset with action potential resolution. |
| Dataset Splits | No | The paper does not explicitly provide training/test/validation splits for any of the datasets used. For the synthetic dataset, it states 'We generated a typical number of 10 trials per neuron.' For the calcium imaging dataset, '15 and 10 repeated trials were recorded, respectively.' For the Neuropixels dataset, it mentions using 'labels of brain area' for evaluation metrics but does not detail how the data was partitioned for model training or evaluation in terms of splits. |
| Hardware Specification | Yes | Computations were performed on an NVIDIA A40 GPU 48 GB. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies such as programming languages, libraries, or frameworks used for the implementation. It mentions "a lightweight multi-layer perceptron (MLP)" but no specific software versions. |
| Experiment Setup | Yes | For the synthetic datasets, we ran all methods with batch size 512 for 100 epochs unless for TS2Vec PCA, where we kept the default batch size and of number of epochs. We trained TRACE and CEED PCA with learning rate 0.3 and TS2Vec with its default learning rate. For the calcium imaging dataset, we used optimal hyperparameters found with a grid search with batch sizes ranging from 1024 to 3200 and learning rates from 0.1 to 0.2 for both TRACE and CEED. The best hyperparameter setting was chosen based on the final loss. For TRACE the best batch size was 1280, while for CEED it was 1024. A learning rate of 0.1 was optimal for both. For the Neuropixels dataset we ran a separate grid search and identified optimal hyperparameters for for TRACE and CEED (0.03 learning rate, batch size 512) and for TRACE + CEED (0.08 learning rate, batch size 1024). Unless otherwise specified, we trained all embeddings for 1000 epochs for the calcium imaging dataset. |