Research map / Label-efficient and robust learning / Self-supervised and contrastive
Self-Supervised Learning with Kernel Dependence Maximization
Yazhe Li, Roman Pogodin, Danica J. Sutherland, Arthur Gretton
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 a target network to SSL-HSIC (ablation) | SSL-HSIC · ImageNet val linear evaluation | Top-1 accuracy: 72.2 -> 74.8 | Improved |
| Add a target network to SSL-HSIC (ablation) | SSL-HSIC · ImageNet val fine-tuning (1% labels) | Top-1 accuracy: 45.3 -> 52.1 | Improved |
| Use SSL-HSIC regularization instead of InfoNCE regularization (ablation) | SSL-HSIC · ImageNet linear evaluation, 100 epochs | Top-1 accuracy: 66.0 -> 66.7 | Improved |
| Use a nonlinear latent kernel rather than a linear kernel (ablation) | SSL-HSIC · ImageNet linear evaluation, 100 epochs | Top-1 accuracy: 65.27 -> 66.67 | Improved |
| Use an IMQ kernel instead of a Gaussian kernel (ablation) | SSL-HSIC · ImageNet linear evaluation, 100 epochs | Top-1 accuracy: 66.67 -> 66.72 | No clear effect |
| Increase the number of random Fourier features (ablation) | SSL-HSIC · ImageNet linear evaluation, 100 epochs | Top-1 accuracy: 66.0 -> 66.4 | Improved |
| Use 512 random Fourier features rather than the exact kernel (ablation) | SSL-HSIC · ImageNet linear evaluation, 100 epochs | Top-1 accuracy: 66.7 -> 66.4 | Hurt |
| Increase training batch size from 256 to 1024 (ablation) | SSL-HSIC · ImageNet linear evaluation, 100 epochs | Top-1 accuracy: 63.7 -> 66.7 | Improved |
| Increase batch size from 1024 to 4096 (ablation) | SSL-HSIC · ImageNet linear evaluation, 100 epochs | Top-1 accuracy: 66.7 -> 66.7 | No clear effect |
| Increase projector/predictor output dimension from 64 to 256 (ablation) | SSL-HSIC · ImageNet linear evaluation, 100 epochs | Top-1 accuracy: 65.4 -> 66.4 | Improved |
| Increase backbone capacity from ResNet-50 (1x) to ResNet-50 (4x) (ablation) | SSL-HSIC · ImageNet linear evaluation | Top-1 accuracy: 74.8 -> 79.1 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Barlow Twins — Barlow Twins: Self-Supervised Learning via Redundancy Reduction (ICML 2021) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Self-Supervised Learning from Structural Invariance (ICLR 2026 · similarity 0.87)
- Improving Self-Supervised Learning by Characterizing Idealized Representations (NeurIPS 2022 · similarity 0.87)
- Making Sense of Dependence: Efficient Black-box Explanations Using Dependence Measure (NeurIPS 2022 · similarity 0.87)
- Mean Shift for Self-Supervised Learning (ICCV 2021 · similarity 0.86)
- Self-supervised Representation Learning from Random Data Projectors (ICLR 2024 · similarity 0.85)
- Non-parametric Representation Learning with Kernels (AAAI 2024 · similarity 0.85)
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.
