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..
Extracting task-relevant preserved dynamics from contrastive aligned neural recordings
Authors: Yiqi Jiang, Kaiwen Sheng, Yujia Gao, Estefany Kelly Buchanan, Yu Shikano, Seung Je Woo, Yixiu Zhao, Tony Hyun Kim, Fatih Dinc, Scott Linderman, Mark J. Schnitzer
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate CANDY on synthetic and real-world datasets spanning multiple species, behaviors, and recording modalities. Our results show that CANDY is able to learn aligned latent embeddings and preserved dynamics across neural recording sessions and subjects, and it achieves improved cross-session behavior decoding performance. We first validated CANDY using a synthetic dataset. Next, we evaluated CANDY s performance on two real-world neural datasets spanning species and recording modalities: two-photon calcium imaging recordings from mouse dorsolateral striatum during a wheel-turning task that we collected and made available with this paper, and publicly available electrophysiological datasets from macaque motor cortex during a center-out reaching task [1, 3, 4]. |
| Researcher Affiliation | Academia | 1Stanford University 2Carnegie Mellon University 3University of California, Santa Barbara 4Howard Hughes Medical Institute EMAIL |
| Pseudocode | No | The paper describes the model architecture and methodology using mathematical equations and diagrams (Fig. 1) but does not include explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code and two-photon imaging data of striatal neural activity that we acquired here are available at https://github.com/schnitzer-lab/CANDY-public.git. |
| Open Datasets | Yes | The code and two-photon imaging data of striatal neural activity that we acquired here are available at https://github.com/schnitzer-lab/CANDY-public.git. and publicly available electrophysiological datasets from macaque motor cortex during a center-out reaching task [1, 3, 4]. The neural and behavioral data were obtained from: https://doi.org/10.48324/dandi.000688/0.250122.1735. |
| Dataset Splits | Yes | To assess both withinand cross-subject generalization, we split data from two mice (24 and 26) to training (60%), validation (10%), and testing (30%) sets, and held out the third mouse (25) entirely for generalization evaluation. For training, we used 40 sessions from this task recorded from subjects C and M. 2 sessions from these two subjects were held out entirely for evaluating the generalization of the pretrained LDS. The ratio of training, validation and testing datasets is 6:1:3. |
| Hardware Specification | Yes | Most experiments performed in this work were performed on Stanford University Sherlock CPU clusters. Running the experiment of contrastive batch size equal to 2048 requires 120GB memory. Some experiments were performed on a desktop with an Intel(R) Core(TM) i9-10900X CPU with 10 physical cores, 256G RAM, and an NVIDIA Ge Force RTX 4070 Ti GPU, or a desktop with an AMD Ryzen 9 9950X 16-core 32-thred CPU, 192G RAM, and an NVIDIA Ge Force RTX 4070 Ti GPU. |
| Software Dependencies | Yes | Neuronal regions of interest (ROIs) were automatically extracted using the EXTRACT algorithm [69] implemented in MATLAB 2022a, followed by manual curation to ensure accurate cell identification. |
| Experiment Setup | Yes | For the mouse wheel-turning dataset, we used Adam optimizer [70] with learning rate of 2 10 3, ℓ2 regularization scale of 5 10 3, encoder and decoder MLP with hidden sizes of [64, 16], 500 training epochs, leaky Re LU activation function, batch size of 32, 2048 time points for contrastive learning, contrastive temperature τ = 0.2, , contrastive scale λcontrastive = 0.1, ℓ1 as the contrastive distance measure, ℓ2 as the contrastive similarity measure, 4-steps-ahead prediction loss, and λbehavior = 0 for Fig. 3 and λbehavior = 1 for Fig. 4 when the behavior supervision loss is enabled. For the monkey center-out reaching dataset, we used Adam optimizer [70] with learning rate of 2 10 3, ℓ2 regularization scale of 2 10 2, encoder and decoder MLP with hidden sizes of [32, 32], 200 training epochs, mish activation function, batch size of 32, 2048 time points for contrastive learning, contrastive temperature τ = 0.5, contrastive scale λcontrastive = 0.1, ℓ1 as the contrastive distance measure, ℓ2 as the contrastive similarity measure, λbehavior = 1, and 2-steps-ahead prediction loss. |