Latent Diffusion for Neural Spiking Data

• DOI: 10.48550/arxiv.2407.08751
• Publication Date: 2024-06-27
• Authors: Jaivardhan Kapoor; Auguste Schulz; Julius Vetter; Felix Pei; Richard Gao; Jakob H. Macke
• Keywords: FOS: Computer and information sciences; Computational Neuroscience; Computer Science - Machine Learning; Quantitative Biology - Neurons and Cognition; FOS: Biological sciences; Neurons and Cognition (q-bio.NC); Data analysis, machine learning and neuroinformatics; Machine Learning (cs.LG)

Modern datasets in neuroscience enable unprecedented inquiries into the relationship between complex behaviors and activity of many simultaneously recorded neurons. While latent variable models can successfully extract low-dimensional embeddings from such recordings, using them to generate realistic spiking data, especially a behavior-dependent manner, still poses challenge. Here, we present Latent Diffusion for Neural Spiking data (LDNS), diffusion-based generative model with space: LDNS employs an autoencoder structured state-space (S4) layers project discrete high-dimensional continuous time-aligned latents. On these inferred latents, train expressive (conditional) diffusion models, enabling us sample neural single-neuron population statistics. We validate on synthetic accurately recovering structure, firing rates, Next, demonstrate its flexibility by generating variable-length that mimics human cortical during attempted speech. show how equip observation accounts dynamics not mediated state, further increasing realism generated samples. Finally, conditional trained motor diverse reaching given reach direction or unseen trajectories. In summary, enables inference latents generation datasets, opening up possibilities simulating experimentally testable hypotheses.