Research map / Probabilistic and scientific modeling / Dynamical systems
Latent Diffusion for Neural Spiking Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LDNS, a two-stage latent diffusion model for neural spiking data using an S4-based autoencoder with Poisson likelihood and a diffusion model over inferred latents, plus post-hoc spike-history observation model, evaluated on Lorenz, human speech BCI, and monkey reach data.
Problem
Generating realistic, behavior-conditioned neural spiking data while also providing low-dimensional latent representations is difficult because spiking data is discrete and existing latent variable models (e.g., VAEs) focus on inference rather than sampling fidelity.
Approach
Compress discrete spike counts into smooth, time-aligned continuous latents with a regularized S4 autoencoder (Poisson NLL, L2 and temporal smoothness, coordinated dropout), then train an S4-based (conditional) DDPM on those latents; add a post-hoc autoregressive spike-history observation model to capture single-neuron dynamics.
Key results
- On monkey reach data, LDNS with spike history (LDNSsh) achieves D_KL of 0.0016 on population spike count histogram versus 0.0036 for AutoLFADSsh.
- LDNS reaches RMSE std ISI 0.023 vs 0.029 for AutoLFADS on monkey reach data (Poisson observation).
- On human BCI, generated variable-length trials match population spike count histograms and single-neuron ISI statistics, with pairwise correlation Pearson r=0.47 to data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add a spike-history autoregressive observation model (ablation) | LDNS · Monkey reach | D_KL population spike count histogram: 0.0039 ± 3.9\text{e-}4 -> 0.0016 ± 6.2\text{e-}4 | Helped |
| add a spike-history autoregressive observation model (ablation) | LDNS · Monkey reach | RMSE pairwise corr: 0.0025 ± 1.1\text{e-}5 -> 0.0025 ± 1.07\text{e-}5 | No clear effect |
| add a spike-history autoregressive observation model (ablation) | LDNS · Monkey reach | RMSE mean isi: 0.037 ± 0.001 -> 0.024 ± 0.002 | Helped |
| add a spike-history autoregressive observation model (ablation) | LDNS · Monkey reach | RMSE std isi: 0.023 ± 0.001 -> 0.023 ± 0.001 | No clear effect |
| add a spike-history autoregressive observation model (ablation) | AutoLFADS · Monkey reach | RMSE mean isi: 0.039 ± 0.003 -> 0.034 ± 0.002 | Helped |
Surprising findings
- Poisson-likelihood-trained LVMs (LFADS and LDNS autoencoder) cannot capture refractoriness-driven autocorrelation dip at 5 ms lags, requiring post-hoc spike-history modeling.
- LFADS sampled over 16x training length showed trajectory divergence and extreme firing rates, while LDNS did not.
Limitations
- Choosing latent dimensionality is challenging and requires prior knowledge
- Autoencoder errors propagate into the latent diffusion model
- LDNS overestimates some pairwise correlations on human BCI data
Experimental setup
Datasets: synthetic Lorenz spiking data human BCI attempted speech dataset (Stanford) MC_Maze monkey reaching dataset
Benchmarks: MC_Maze monkey reaching human BCI attempted speech synthetic Lorenz
Metrics: KL divergence of population spike count histogram RMSE of pairwise correlations RMSE of mean ISI RMSE of std ISI power spectral density Pearson correlation
Baselines: AutoLFADS LFADS with spike history TNDM pi-VAE
Builds on: latent diffusion models denoising diffusion probabilistic models S4 LFADS coordinated dropout
Closest papers by meaning
- Learning Delays in Spiking Neural Networks using Dilated Convolutions with Learnable Spacings (ICLR 2024 · similarity 0.88)
- P-SPIKESSM: HARNESSING PROBABILISTIC SPIKING STATE SPACE MODELS FOR LONG-RANGE DEPENDENCY TASKS (ICLR 2025 · similarity 0.88)
- A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks (NeurIPS 2025 · similarity 0.87)
- NDOT: Neuronal Dynamics-based Online Training for Spiking Neural Networks (ICML 2024 · similarity 0.87)
- Causal pieces: analysing and improving spiking neural networks piece by piece (NeurIPS 2026 · similarity 0.87)
- TS-LIF: A Temporal Segment Spiking Neuron Network for Time Series Forecasting (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.