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..
Know Thyself by Knowing Others: Learning Neuron Identity from Population Context
Authors: Vinam Arora, Divyansha Lachi, Ian Knight, Mehdi Azabou, Blake A. Richards, Cole Hurwitz, Joshua H Siegle, Eva L Dyer
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Across multiple electrophysiology and calcium imaging datasets, a linear decoding evaluation on top of Nu CLR representations achieves a new state-of-the-art for both cell type and brain region decoding tasks, and demonstrates strong zero-shot generalization to unseen animals. |
| Researcher Affiliation | Academia | 1University of Pennsylvania, 2Columbia University, 3Mc Gill University, 4Mila, 5Allen Institute for Neural Dynamics |
| Pseudocode | No | The paper describes the model architecture and training objective in Sections 2.1 and 2.2 with mathematical equations and descriptive text, but no explicit pseudocode blocks are present. |
| Open Source Code | Yes | Code is available at https://github.com/nerdslab/nuclr. |
| Open Datasets | Yes | On the Allen Visual Coding Neuropixels dataset [41], doubling the unlabeled pretraining data improves cell-type decoding more than doubling the supervised labels, highlighting the advantages of scaling pretraining across many animals. ... The second dataset that we evaluate on is a spatial transcriptomics dataset from Bugeon et. al. [7], which consists of calcium imaging recordings from 17 sessions across 4 mice... For brain region identification, we evaluate on two electrophysiological datasets with well-curated anatomical annotations: the International Brain Laboratory (IBL) Brain-wide Map [2] and the Steinmetz et al. 2019 dataset [43]. |
| Dataset Splits | Yes | For each downstream task, we perform evaluation across three generalization settings: (1) Transductive, where the testing populations are seen during self-supervised pretraining, and partial labels from these populations are used to train the classification head; (2) Transductive zero-shot, where the test populations are present during pretraining, but no labels from them are used when training the classifier; and (3) Inductive zero-shot, where the test populations are entirely unseen during pretraining, and no fine-tuning is performed on the encoder or classifier before evaluation. ... In total, 93 recordings are designated for the test set, and the remaining 346 are used for training and validation. Of the training set, 91 recordings are used as a development set for tuning hyperparameters specific to electrophysiology data. ... For non-zero-shot evaluation, we create a neuron-wise stratified test-train split with a test size of 20%. ... For transductive zero-shot evaluation, we use a 10-fold leave-one-subject-out strategy. In each fold, one subject is held out for testing, and the remaining nine are used for training. |
| Hardware Specification | Yes | Pretraining takes approximately 3 hours on a machine with 4 NVidia H100 GPUs. |
| Software Dependencies | No | We use standard scaled-dot-product attention as implemented in xformers [25]... The first view is sampled using the Random Fixed Window Sampler found in torch_brain9. |
| Experiment Setup | Yes | All relevant hyperparameters for training Nu CLR are listed in Table 9. These values were mainly selected via manual line searches on the IBL development set (for ephys data, Appendix D.2) and the Bugeon et al. development set (for calcium imaging data, Appendix D.4). Across all datsets, we pretrain the model for 50,000 steps and use the bfloat16 number format throughout. ... We use the Adam W [28] optimizer with a linear learning rate warm-up over the first epoch, followed by cosine decay until end of training. |