Research map / Probabilistic and scientific modeling / Variational and generative sampling
SPLICE: Fully Tractable Hierarchical Extension of ICA with Pooling
Jun-ichiro Hirayama, Aapo Hyvärinen, Motoaki Kawanabe
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ML fine-tuning on top of layerwise SPLICE-LW (SPLICE-ML) (ablation) | SPLICE two-layer · Natural Images, n=200,000 | test avg. log-likelihood: -221.18 -> -199.67 | Improved |
| Replace single-layer ICA with two-layer SPLICE-LW | ICA · Natural Images, n=100,000 | test avg. log-likelihood: - 270.09 -> - 222.21 | Improved |
| Replace EBM-NCE with SPLICE-ML | hierarchical ICA with pooling · Natural Images, n=200,000 | test avg. log-likelihood: -330.04 -> -199.67 | Improved |
| Evaluate partition function numerically by AIS instead of analytically (ablation) | SPLICE-ML · Natural Images, n=100,000 | test avg. log-likelihood: - 200.77 -> -200.76 | No clear effect |
| Increase training sample size from 100,000 to 200,000 | SPLICE-ML · Natural Images | test avg. log-likelihood: - 200.77 -> -199.67 | No clear effect |
| Increase sample size from 100,000 to 200,000 | EBM-NCE · Natural Images | test avg. log-likelihood: -342.12 -> -330.04 | Improved |
| Increase number of layers from 2 to 4 (extra pooling and ICA layers) (ablation) | SPLICE-LW · Natural Images, n=200,000 | test avg. log-likelihood | Improved |
| Increase number of layers from 2 to 4 (ablation) | SPLICE-ML · Natural Images, n=200,000 | test avg. log-likelihood | No clear effect |
| Use SPLICE instead of EBM-NCE for second-layer source separation | two-layer model · Synthetic complex data, Layer 2 | performance index / mean abs. correlation | Improved |
| Use SPLICE instead of EBM/ICA-style estimation for first layer | two-layer model · Synthetic complex data, Layer 1, small sample sizes | performance index | Improved |
| Add unsupervised ML fine-tuning to SPLICE-LW (ablation) | SPLICE · Synthetic data | performance index | Improved |
| Use SPLICE-LW/ML instead of EBM-NCE for second-layer sources | two-layer model · EEG BCI IV dataset 1 (original and transfer) | AUC of second-layer components for task state | Improved |
Closest papers by meaning
- A Normative and Biologically Plausible Algorithm for Independent Component Analysis (NeurIPS 2021 · similarity 0.87)
- On the Identifiability of Nonlinear ICA: Sparsity and Beyond (NeurIPS 2022 · similarity 0.87)
- Shared Independent Component Analysis for Multi-Subject Neuroimaging (NeurIPS 2021 · similarity 0.86)
- Modeling Shared responses in Neuroimaging Studies through MultiView ICA (NeurIPS 2020 · similarity 0.86)
- On the Identifiability of Sparse ICA without Assuming Non-Gaussianity (NeurIPS 2023 · similarity 0.86)
- Unconstrained Stochastic CCA: Unifying Multiview and Self-Supervised Learning (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
