Research map / Label-efficient and robust learning / Recognition and long-tail
An Empirical Study on Disentanglement of Negative-free Contrastive Learning
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 |
|---|---|---|---|
| Evaluate only the top-k most disentangled dimensions per factor instead of the full representation (ablation) | BYOL · dSprites | MED: 31.3 (0.4) -> 53.7 (0.7) | Helped |
| Evaluate only the top-k most disentangled dimensions per factor instead of the full representation (ablation) | BYOL · Cars3D | MED: 9.7 (0.5) -> 31.8 (1.3) | Helped |
| Increase representation dimension (ablation) | BYOL · dSprites | MED and Top-2 MED | Helped |
| Increase representation dimension (ablation) | VAE-based methods · dSprites | MED and Top-k MED | Hurt |
| Use negative-free contrastive learning rather than the strongest listed low-dimensional generative baseline | contrastive methods · Shapes3D | MED | Hurt |
| Use negative-free contrastive learning rather than the strongest listed low-dimensional generative baseline | contrastive methods · Cars3D | MED | Hurt |
| Use negative-free contrastive learning rather than the strongest listed low-dimensional generative baseline | contrastive methods · SmallNORB | MED | Hurt |
| Use negative-free contrastive learning rather than the strongest listed low-dimensional generative baseline | contrastive methods · CelebA | MED | Helped |
| Extract a selected subspace from high-dimensional contrastive representations instead of requiring the full representation to be disentangled | negative-free contrastive methods · synthetic datasets and CelebA | Top-k MED | Helped |
| Evaluate a 1000-d BYOL representation with MED rather than DCI using Gradient Boosting Tree | BYOL on Cars3D · Cars3D | evaluation time: more than 14 hours -> less than 20 seconds | Helped |
| Use existing disentanglement metrics on high-dimensional representations rather than MED | high-dimensional BYOL · SmallNORB | disentanglement measurement agreement | Hurt |
Closest papers by meaning
- Towards a Unified Framework of Contrastive Learning for Disentangled Representations (NeurIPS 2023 · similarity 0.92)
- Theory and Evaluation Metrics for Learning Disentangled Representations (ICLR 2020 · similarity 0.90)
- Disentangled Representation Learning via Flow Matching (NeurIPS 2026 · similarity 0.88)
- Non-negative Contrastive Learning (ICLR 2024 · similarity 0.87)
- Enriching Disentanglement: From Logical Definitions to Quantitative Metrics (NeurIPS 2024 · similarity 0.87)
- Graph-based Unsupervised Disentangled Representation Learning via Multimodal Large Language Models (NeurIPS 2024 · similarity 0.87)
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.