Disentangling Identifiable Features from Noisy Data with Structured Nonlinear ICA
Authors: Hermanni Hälvä, Sylvain Le Corff, Luc Lehéricy, Jonathan So, Yongjie Zhu, Elisabeth Gassiat, Aapo Hyvarinen
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments: To explore this, we first performed experiments on simulated data. We compared the performance of our model to the current state-of-the-art, IIA-HMM (Morioka et al., 2021), as well as identifiable VAE (i VAE) (Khemakhem et al., 2020a) and standard linear Gaussian state-space model (LGSSM). (...) We simulated 100K long time-sequences from the SNICA model and computed the mean absolute correlation coefficient (MCC) between the estimated latent components and ground truth independent components (...). Application to denoising -SNICA is able to denoise time-series signals by learning the generative model and then performing inference on latent variables. (...) 5.2 Experiments on real MEG data: To demonstrate real-data applicability, -SNICA was applied to multivariate time series of electrical activity in the human brain, measured by magnetoencephalography (MEG). (...) Figure 3 a) shows the classification accuracies of the stimulus categories, across different methods and the number of layers for each model. |
| Researcher Affiliation | Academia | 1Department of Computer Science, University of Helsinki, Finland 2 Samovar, Télécom Sud Paris, département CITI, Institut Polytechnique de Paris, Palaiseau, France 3Laboratoire J. A. Dieudonné, Université Côte d Azur, CNRS, 06100, Nice, France 4Department of Engineering, University of Cambridge, UK 5Université Paris-Saclay, CNRS, Laboratoire de mathématiques d Orsay, 91405, Orsay, France |
| Pseudocode | No | The paper describes the estimation method using narrative text and mathematical equations in Section 5 and Appendix C, but it does not provide any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code will be openly available at https://github.com/HHalva/snica. |
| Open Datasets | Yes | We considered a resting state MEG sessions from the Cam-CAN dataset. During the resting state recording, subjects sat still with their eyes closed. |
| Dataset Splits | Yes | The performance was evaluated by the generalizability of a classifier across subjects. i.e., one-subject-out cross-validation. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments. It does not mention any CPU, GPU, or specific cloud instances. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies. While it mentions code will be available on GitHub, the text itself does not list software names with versions. |
| Experiment Setup | Yes | We simulated 100K long time-sequences from the SNICA model (...). More precisely, to illustrate the dimensionality reduction capabilities we considered two settings where the observed data dimension M, was either 12 or 24 and the number of independent components, N was 3 and 6, respectively. (...) We considered four levels of mixing of increasing complexity by randomly initialized MLPs of the following number of layers: 1 (linear ICA), 2, 3, and 5. (...) The number of layers of the decoder and encoder were equal and took values 2, 3, 4. We fixed the number of independent components to 5 so that our result can be fairly compared to those in Morioka et al. (2021). |