nerds_lab

About the NerDS Lab

The Neural Data Science (NerDS) Lab develops foundation models and representation learning methods for neuroscience, behavior, and biological systems. Our research combines machine learning, neural data analysis, and AI for science to build scalable models that decode neural computation, learn transferable representations across animals and recording modalities, and uncover general principles of biological intelligence.

We work on neural foundation models, self-supervised learning, neural forecasting, graph representation learning, and multimodal AI systems that integrate behavior, neural activity, and biological structure across scales.

🧠 Neural Foundation Models We develop large-scale foundation models for neural activity that learn transferable representations across animals, brain regions, recording modalities, and behavioral tasks. Our work studies how large-scale pretraining enables zero-shot decoding, neural forecasting, and the discovery of shared computational principles across neural systems. Featured Papers: A unified, scalable framework for neural population decoding (NeurIPS 2023) Multi-session, multi-task neural decoding from distinct cell-types and brain regions (ICLR 2025) Towards a universal translator for neural dynamics at single-cell, single-spike resolution (NeurIPS 2024) Generalizable, real-time neural decoding with hybrid state-space models (NeurIPS 2025)

🧬 Neuron Identity and Brain Organization We build machine learning systems that reveal structure in large-scale neural recordings, including representations of neuron identity, brain regions, cell types, and population dynamics. Our models learn from neural activity directly and generalize across animals, sessions, and experimental conditions. Featured Papers: Know Thyself by Knowing Others: Learning Neuron Identity from Population Context (NeurIPS 2025) Multi-session, multi-task neural decoding from distinct cell-types and brain regions (ICLR 2025) Transcriptomic cell type structures in vivo neuronal activity across multiple timescales (Cell Reports 2023)

🔁 Self-Supervised and Contrastive Learning We develop self-supervised and contrastive learning methods that uncover structure in neural, behavioral, and biological data without requiring labels. Our work focuses on multiscale representation learning, event-driven objectives, and transfer across datasets, tasks, and modalities. Featured Papers: Relax, it doesn't matter how you get there (NeurIPS 2023 Spotlight) Your contrastive learning problem is secretly a distribution alignment problem (NeurIPS 2024) Drop, Swap, and Generate (NeurIPS 2021 Oral)

📈 Neural Forecasting and Generative Modeling We develop forecasting and generative models that predict neural and behavioral dynamics across multiple timescales. Our work studies how structured temporal representations, event-driven learning, and multiscale dynamics improve forecasting, simulation, and representation learning in biological systems. Featured Papers: PRISM: A Hierarchical Multiscale Approach for Time Series Forecasting (2026) Towards a universal translator for neural dynamics at single-cell, single-spike resolution (NeurIPS 2024) Drop, Swap, and Generate (NeurIPS 2021 Oral)

🌍 Domain Adaptation and Representation Alignment We design representation alignment methods that adapt across subjects, devices, datasets, and recording conditions. Our work uses optimal transport, distribution alignment, and domain generalization to build models that remain robust under real-world shifts in neural, behavioral, and physiological time series. Featured Papers: Your contrastive learning problem is secretly a distribution alignment problem (NeurIPS 2024) Time Series Domain Adaptation via Channel-Selective Representation Alignment (TMLR 2025) Aligning latent representations of neural activity (Nature Biomedical Engineering 2022) Making transport more robust and interpretable by moving data through anchor points (ICML 2021)

📊 Graph Learning We develop graph transformers, relational learning methods, and graph foundation models that learn transferable representations across diverse graph domains. Our work studies scalable graph pretraining, multimodal relational learning, and representation transfer across heterogeneous graph structures. Featured Papers: GraphFM: A generalist graph transformer that learns transferable representations across diverse domains (TMLR 2025) RGP: A Cross-Attention based Graph Transformer for Relational Deep Learning (LOG 2025) Half-Hop: A graph upsampling approach for slowing down message passing (ICML 2023)

Funding

We are fortunate to receive funding from the following sources: