Research map / Probabilistic and scientific modeling / Causal discovery
Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use InfoNCE on augmented views while style variables may change | Synthetic latent representation learning · Synthetic settings with statistical and causal dependencies | R² (Content): 0.98 ± 0.03; 0.99 ± 0.01 | Helped |
| Introduce statistical dependence within content and style | Synthetic InfoNCE encoder · Synthetic, p(chg.) 0.75 | R² (Style): 0.06 ± 0.05 -> 0.37 ± 0.05 | Helped |
| Add causal dependence of style on content | Synthetic InfoNCE encoder with statistical dependence · Synthetic, p(chg.) 0.75 | R² (Style): 0.37 ± 0.05 -> 0.80 ± 0.08 | Helped |
| Apply colour distortion to views | Causal3DIdent InfoNCE encoder · Causal3DIdent | R² (object hue): 0.17 ± 0.00 | Hurt |
| Apply colour distortion to views | Causal3DIdent InfoNCE encoder · Causal3DIdent | R² (object position): 0.61 ± 0.10 | Helped |
| Use large crops | Causal3DIdent InfoNCE encoder · Causal3DIdent | R² (object hue): 0.87 ± 0.00 | Helped |
| Use small instead of large crops | Causal3DIdent InfoNCE encoder · Causal3DIdent | R² (object hue): 0.87 ± 0.00 -> 0.00 ± 0.01 | Hurt |
| Combine crop and colour distortion | Causal3DIdent InfoNCE encoder · Causal3DIdent | R² (class): 0.42 ± 0.01 -> 1.00 ± 0.00 | Helped |
| Apply image-level rotation | Causal3DIdent InfoNCE encoder · Causal3DIdent | R² (object rotation): 0.05 ± 0.03 | Hurt |
| Set representation dimension below content dimension (ablation) | Synthetic content representation · Synthetic setting | Content information retained | Hurt |
| Set representation dimension above content dimension (ablation) | Synthetic content representation · Synthetic setting | Style information encoded | Helped |
| Use BarlowTwins instead of SimCLR (ablation) | Synthetic representation learning experiment · Synthetic setting | R² | No clear effect |
Closest papers by meaning
- Content-Style Learning from Unaligned Domains: Identifiability under Unknown Latent Dimensions (ICLR 2025 · similarity 0.87)
- Understanding Augmentation-based Self-Supervised Representation Learning via RKHS Approximation and Regression (ICLR 2024 · similarity 0.87)
- Multi-View Causal Representation Learning with Partial Observability (ICLR 2024 · similarity 0.86)
- Learning Discrete Concepts in Latent Hierarchical Models (NeurIPS 2024 · similarity 0.86)
- An Augmentation-Aware Theory for Self-Supervised Contrastive Learning (ICML 2025 · similarity 0.86)
- Object centric architectures enable efficient causal representation learning (ICLR 2024 · similarity 0.86)
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.