Reproduce “Latent Diffusion for Neural Spiking Data” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jaivardhan Kapoor, Auguste Schulz, Julius Vetter, Felix Pei, Richard Gao, Jakob H. Macke
LDNS compresses binned spike counts into latents, trains a diffusion model to generate them, and decodes them into rates and spikes, optionally conditioned on behavior. In held-out monkey reach tests, generated activity decoded to movements consistent with supplied trajectories, not exact copies. With spike-history observations, LDNS’s spike-count-histogram KL divergence was 0.0016 versus 0.0036 for AutoLFADS; this comparison covers one metric, not every measure.
Key insights
- Denoising diffusion models, when applied to a continuous latent representation of neural activity, can effectively generate high-fidelity and diverse samples of discrete spiking data.
- Structured State-Space (S4) layers are crucial for handling variable-length neural recordings and enabling robust length generalization in both latent inference and generation.
- A modular architecture, separating latent encoding/decoding from the generative process, allows for flexible conditional generation and the incorporation of specific neural dynamics like spike history.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.