SPLICE: Fully Tractable Hierarchical Extension of ICA with Pooling
Authors: Jun-ichiro Hirayama, Aapo Hyvärinen, Motoaki Kawanabe
ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 4. Experimental Results: In this section, we demonstrate our SPLICE in a simulation study and with two motivating real datasets. |
| Researcher Affiliation | Academia | 1RIKEN Center for Advanced Intelligence Project (AIP), Tokyo, Japan 2Advanced Telecommunications Research Institute International (ATR), Kyoto, Japan 3Department of Computer Science and HIIT, University of Helsinki, Finland 4Gatsby Computational Neuroscience Unit, University College London, UK. |
| Pseudocode | No | The paper describes methods textually and mathematically but does not include any structured pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper does not provide any concrete access to source code for the described methodology (e.g., no repository link, explicit code release statement, or mention of code in supplementary materials). |
| Open Datasets | Yes | To further demonstrate the applicability to exploratory analysis of neuroimaging signals, we applied the same methods to a publicly available EEG datasets: Datasets 1 (Blankertz et al., 2007) from the BCI competition IV (http://www.bbci.de/competition/iv/). natural images obtained from Image Net10K (Deng et al., 2010). |
| Dataset Splits | Yes | We first quantitatively compared two-layer SPLICE and EBM (as well as a single-layer ICA) by their test log-likelihood evaluated with 10-fold cross-validation (CV) |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper mentions using 'Matlab code' from a GitHub repository for Hamiltonian-Annealed-Importance-Sampling and refers to 'Fast ICA', but it does not provide specific version numbers for these or any other key software components (e.g., 'Matlab 2018a' or 'Fast ICA v1.2'). |
| Experiment Setup | Yes | The dataset consisted of 30-dimensional complex-valued vectors xt which we synthesized by our generative model. We generated the top-layer sources s k from t-distribution of the three degrees of freedom and every entry in A and (the real and complex parts of) A uniformly in [-1, 1]. For simplicity, we used the same Gaussianization-based F 1 to generate the data. Image patches were of 32x32 pixels, with the pixel values in each patch normalized to have zero mean and unit variance. The dimensionality was then reduced to d = 200 by PCA. The logarithmic nonlinearity F was commonly used in both SPLICE and EBM. The number of second-layer sources was commonly set as m = 50. |