Latent Diffusion for Neural Spiking Data
Authors: Jaivardhan Kapoor, Auguste Schulz, Julius Vetter, Felix Pei, Richard Gao, Jakob H Macke
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate LDNS on synthetic data, accurately recovering latent structure, firing rates, and spiking statistics. Next, we demonstrate its flexibility by generating variable-length data that mimics human cortical activity during attempted speech. We show how to equip LDNS with an expressive observation model that accounts for single-neuron dynamics not mediated by the latent state, further increasing the realism of generated samples. Finally, conditional LDNS trained on motor cortical activity during diverse reaching behaviors can generate realistic spiking data given reach direction or unseen reach trajectories. |
| Researcher Affiliation | Academia | 1Machine Learning in Science, University of Tübingen & Tübingen AI Center, Tübingen, Germany 2Department Empirical Inference, Max Planck Institute for Intelligent Systems, Tübingen, Germany |
| Pseudocode | No | The paper describes algorithms and network architectures using textual descriptions and diagrams (Figure A1), but it does not include formal pseudocode blocks or sections explicitly labeled 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | Code available at https://github.com/mackelab/LDNS. |
| Open Datasets | Yes | All real-world datasets used in this work are publicly available under open-access licenses. The human BCI dataset is available at https://datadryad.org/stash/downloads/file_ stream/2547369 under a CC0 1.0 Universal Public Domain Dedication license. This dataset was originally published in Willett et al. [57]. The monkey reaching dataset (MC_Maze) is available through the DANDI Archive (https:// dandiarchive.org/dandiset/000128, ID: 000128) under a CC-BY-4.0 license. |
| Dataset Splits | No | The paper mentions 'Num training trials' in Table 2 and evaluating on a 'test set' for synthetic data but does not provide explicit training, validation, or test dataset split percentages or counts for all datasets. |
| Hardware Specification | Yes | We performed all training and evaluation of LDNS on the Lorenz and Monkey reach datasets on an NVIDIA RTX 3090 GPU with 24GB RAM. For the Human BCI data, we used an NVIDIA A100 40GB GPU. Auto LFADS, the baseline used for unconditional sampling for the Monkey reach dataset, was trained on a cluster of 8 NVIDIA RTX 2080TI GPUs for one day. |
| Software Dependencies | No | The paper mentions using 'Pytorch framework' and 'Weights & Biases' but does not specify their version numbers. It also refers to an 'S4 implementation' without a version. |
| Experiment Setup | Yes | Table 2: Training details for autoencoder models on Lorenz, Monkey reach, and Human BCI datasets. We used the Adam W [31] optimizer, whose learning rate was linearly increased over in the initial period and then decayed to 10% of the max value with a cosine schedule. Mean firing rate for Lorenz was 0.3. In all cases, we used K = 5 for the temporal smoothness loss in Eq. 1. (Parameter examples: Max learning rate 0.001, Adam W weight decay 0.01, Batch size 512, L2 reg. β1 0.01) Table 3: Training details for diffusion models on Lorenz, Monkey reach, and Human BCI datasets. We used the same learning rate scheduler as for the autoencoder. (Parameter examples: Num diffusion blocks 4, Num denoising steps 1000, Max learning rate 0.001, Batch size 512) |