Manifold GPLVMs for discovering non-Euclidean latent structure in neural data
Authors: Kristopher Jensen, Ta-Chu Kao, Marco Tripodi, Guillaume Hennequin
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the validity of the approach on several synthetic datasets, as well as on calcium recordings from the ellipsoid body of Drosophila melanogaster and extracellular recordings from the mouse anterodorsal thalamic nucleus. |
| Researcher Affiliation | Academia | Kristopher T. Jensen Department of Engineering University of Cambridge ktj21@cam.ac.uk Ta-Chu Kao Department of Engineering University of Cambridge tck29@cam.ac.uk Marco Tripodi Neurobiology Division MRC Laboratory of Molecular Biology mtripodi@mrc-lmb.cam.ac.uk Guillaume Hennequin Department of Engineering University of Cambridge g.hennequin@eng.cam.ac.uk |
| Pseudocode | No | The paper describes the generative model, inference process, and mathematical formulations, but it does not include any explicit pseudocode blocks or algorithms. |
| Open Source Code | No | The paper discusses the broader impact of methods for probabilistic inference and mentions open-source code as a general concept in the 'Broader Impact' section, but it does not provide an explicit statement or link to the source code for the methodology described in this paper. |
| Open Datasets | Yes | Finally, we apply m GPLVM to a biological dataset to show that it is robust to the noise and heterogeneity characteristic of experimental recordings. Here, we used calcium imaging data recorded from the ellipsoid body (EB) of Drosophila melanogaster (Turner-Evans, 2020; Turner-Evans et al., 2020), where the so-called E-PG neurons have recently been shown to encode head direction (Seelig and Jayaraman, 2015). The EB is divided into 16 wedges , each containing 2-3 E-PG neurons that are not distinguishable on the basis of calcium imaging data, and we therefore treat each wedge as one neuron . Due to the physical shape of the EB, neurons come pre-ordered since their joint activity resembles a bump rotating on a ring (Figure 4a, analogous to Figure 2, ordered data ). In Appendix A, we describe similar results with m GPLVM applied to a dataset from the mouse head-direction circuit with more heterogeneous neuronal tuning and no obvious anatomical organization (Peyrache et al., 2015). Turner-Evans, D. B. (2020). Kir.zip. Janelia Research Campus. Dataset. https://doi.org/10.25378/janelia.12490325.v1. Peyrache, A., Petersen, P., and Buzsáki, G. (2015). Extracellular recordings from multi-site silicon probes in the anterior thalamus and subicular formation of freely moving mice. CRCNS.org. Dataset. https://doi.org/10.6080/K0G15XS1. |
| Dataset Splits | Yes | We simulated 10 different toroidal datasets; for each, we used half the conditions to fit the GP hyperparameters, and half the neurons to predict the latent states for the conditions not used to fit the GP parameters. Finally, we used the inferred GP parameters and latent states to predict the activity of the held-out neurons at the held-out conditions. As expected, the predictions of the toroidal model outperformed those of the standard Euclidean GPLVM which cannot capture the periodic boundary conditions of the torus (Figure 3c). We fitted the full m GPLVM with a separate GP for each neuron and found that T 1-m GPLVM performed better than R1-m GPLVM on both cross-validated prediction errors and log marginal likelihoods (Figure 4b). Each datapoint corresponds to a different partition of the timepoints into a training set and a test set. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware (e.g., CPU, GPU models, memory) used to run the experiments. |
| Software Dependencies | No | The paper describes the mathematical model and inference methods (e.g., variational inference, reparameterization trick, sparse GP approximation) but does not list any specific software components or libraries with their version numbers that were used for implementation. |
| Experiment Setup | Yes | For computational simplicity, here we constrain the m GPLVM parameters αi, ℓi and σi to be identical across neurons. We simulated 10 different toroidal datasets; for each, we used half the conditions to fit the GP hyperparameters, and half the neurons to predict the latent states for the conditions not used to fit the GP parameters. Finally, we used the inferred GP parameters and latent states to predict the activity of the held-out neurons at the held-out conditions. |