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..
POCO: Scalable Neural Forecasting through Population Conditioning
Authors: Yu Duan, Hamza Chaudhry, Misha B Ahrens, Christopher Harvey, Matthew G Perich, Karl Deisseroth, Kanaka Rajan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Trained across five calcium imaging datasets spanning zebrafish, mice, and C. elegans, POCO achieves state-of-the-art accuracy at cellular resolution in spontaneous behaviors. After pre-training, POCO rapidly adapts to new recordings with minimal fine-tuning. Notably, POCO s learned unit embeddings recover biologically meaningful structure such as brain region clustering without any anatomical labels. Our comprehensive analysis reveals several key factors influencing performance, including context length, session diversity, and preprocessing. |
| Researcher Affiliation | Academia | Yu Duan1,5 Hamza Tahir Chaudhry2 Misha B. Ahrens 3 Christopher D. Harvey4 Matthew G. Perich 6,7 Karl Deisseroth8 Kanaka Rajan4,5 1EECS, MIT 2SEAS, Harvard University 3Janelia Research Campus, HHMI 4 Harvard Medical School 5Kempner Institute 6Université de Montréal 7 Mila Quebec AI Institute 8 Stanford University Corresponding Authors EMAIL, EMAIL |
| Pseudocode | No | The paper includes architectural diagrams (Figure 1) and mathematical equations describing the model, but no explicit 'Pseudocode' or 'Algorithm' blocks are provided. |
| Open Source Code | Yes | Code is available at https://github.com/yuvenduan/POCO. |
| Open Datasets | Yes | Three datasets (Ahrens zebrafish, Zimmer and Flavell C. elegans) are publicly available; the remaining two (Deisseroth zebrafish and Harvey mouse) are available upon request or under collaboration agreements. |
| Dataset Splits | Yes | We first cut each session into 1K-step segments, then partitioned each segment into training, validation, and test sets by 3:1:1. |
| Hardware Specification | Yes | All models are trained on NVIDIA A100 or H100 GPUs paired with AMD EPYC 9454 CPUs. |
| Software Dependencies | No | The paper mentions using Adam W optimizer but does not specify software libraries with version numbers (e.g., PyTorch 1.x, TensorFlow 2.x, Python 3.x, CUDA 11.x). |
| Experiment Setup | Yes | For all models, we used Adam W [30] optimizer with learning rate 0.0003 and weight decay 10 4. At each training step, a batch of 64 sequences is sampled from the training sets of all sessions. For single-cell prediction in larval zebrafish, we reduce the batch size to 8 for Deisseroth s zebrafish dataset and to 4 for Ahrens zebrafish dataset due to memory constraints. Similarly, for the MLP conditioned by univariate Transformer, we reduce the batch size to 16. Gradient clipping is applied to limit the gradient norm to 5. Most models are trained using mean squared error (MSE) loss averaged over the P prediction steps. |