Research map / Label-efficient and robust learning / Self-supervised and contrastive
Learning Interpretable Low-dimensional Representation via Physical Symmetry
Xuanjie Liu, Daniel Chin, Yichen Huang, Gus Xia
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 latent-space physical symmetry augmentation (K=4) (ablation) | SPS_VAE for monophonic music · Monophonic music test set | Synthesis R²: 0.910$\pm$0.040 -> \textbf{0.986$\pm$0.025} | Improved |
| Add latent-space physical symmetry augmentation (K=4) (ablation) | SPS_AE for monophonic music · Monophonic music test set | Synthesis R²: 0.832$\pm$0.129 -> \textbf{0.986$\pm$0.025} | Improved |
| Add latent-space physical symmetry augmentation (K=4) (ablation) | SPS_VAE for bouncing-ball videos · Bouncing ball test set | Location linear-fit MSE: 0.62 $\pm$ 0.14 -> 0.26 $\pm$ 0.30 | Improved |
| Add latent-space physical symmetry augmentation (K=4) (ablation) | SPS_AE for bouncing-ball videos · Bouncing ball test set | Location linear-fit MSE: 0.42 $\pm$ 0.15 -> \textbf{0.24 $\pm$ 0.17} | Improved |
| Use an AE instead of a VAE | SPS for monophonic music · Monophonic music test set | Synthesis R²: 0.986$\pm$0.025 -> \textbf{0.986$\pm$0.025} | No clear effect |
| Use an AE instead of a VAE | SPS for bouncing-ball videos · Bouncing ball test set | Location linear-fit MSE: 0.26 $\pm$ 0.30 -> \textbf{0.24 $\pm$ 0.17} | No clear effect |
| Replace SPS with reconstruction-only β-VAE | Monophonic music representation learning · Monophonic music test set | Learned factor R²: 0.999$\pm$0.001 -> 0.772$\pm$0.333 | Hurt |
| Remove the separate style representation (ablation) | SPS_VAE+ for multiple-instrument audio · Multiple-instrument audio test set | Timbre augmentation precision: 0.98 -> 0.50 | Hurt |
| Remove the separate style representation (ablation) | SPS_VAE+ for colourful bouncing-ball videos · Colourful bouncing ball test set | Colour augmentation precision: 0.99 -> 0.64 | Hurt |
| Increase the counterfactual augmentation factor K | SPS for bouncing-ball videos · Bouncing ball training sets of 256, 512, 1024, and 2048 samples | Linear projection loss | Improved |
| Use perturbed, deliberately incorrect symmetry groups instead of no symmetry | SPS for bouncing-ball videos · Bouncing ball dataset | Linear projection loss | Improved |
| Add symmetry constraints to the sequential autoencoder (ablation) | SPS for audio · Monophonic music test set | Prior prediction loss: 0.0030$\pm$0.0033 -> 0.0012$\pm$0.0006 | Improved |
Closest papers by meaning
- From Isolated feature to Orbits:\\Discovering Music Concepts via Multi-SAE Alignment (NeurIPS 2026 · similarity 0.87)
- Unsupervised Disentanglement of Content and Style via Variance-Invariance Constraints (ICLR 2025 · similarity 0.84)
- Contrastive Learning from Synthetic Audio Doppelgängers (ICLR 2025 · similarity 0.84)
- MeLFusion: Synthesizing Music from Image and Language Cues using Diffusion Models (CVPR 2024 · similarity 0.83)
- MoMu-Diffusion: On Learning Long-Term Motion-Music Synchronization and Correspondence (NeurIPS 2024 · similarity 0.83)
- MusER: Musical Element-Based Regularization for Generating Symbolic Music with Emotion (AAAI 2024 · similarity 0.83)
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.
