Research map / Label-efficient and robust learning / Self-supervised and contrastive
Enhancing Contrastive Learning Inspired by the Philosophy of “The Blind Men and the Elephant”
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes JointCrop and JointBlur, which sample the two augmentation parameters jointly (via controlling area or blur-kernel ratios) so contrastive positive pairs are more challenging, yielding gains for several SSL methods at no extra cost.
Problem
Standard contrastive learning samples the two augmentation parameters of a positive pair independently, which yields many trivial or incomprehensive positive pairs (e.g., 80% of RandomCrop pairs have area ratios within 1:2).
Approach
Instead of sampling the two augmentation parameters independently, first sample their ratio (crop-area ratio or blur-sigma ratio) from a controllable distribution JC(beta) and then sample the first parameter conditioned on it, so the joint distribution is non-product and yields harder positive pairs; the approach is plug-and-play with no extra compute.
Key results
- JointCrop(0) improves MoCo v1 on ImageNet-1K linear eval from 60.6 to 63.29 (+2.69) and SimCLR from 60.7 to 62.16 (+1.46) at 100 epochs.
- On Tiny-ImageNet, J-Crop(-2) improves SimCLR from 45.25 to 47.95 and BYOL from 48.91 to 51.02.
- JointCrop adds negligible training cost: MoCo v1 788.2 s vs 782.8 s with JointCrop over 5 epochs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add JointCrop (β=0) | SimCLR · ImageNet-1K | top-1 accuracy: 60.7 -> 62.16 (+1.46) | Helped |
| Add JointBlur (β=0) | MoCo v1 · ImageNet-1K | top-1 accuracy: 60.6 -> 62.87 (+2.27) | Helped |
| Add JointCrop (β=0) | MoCo v3 · ImageNet-1K, 100 epochs | top-1 accuracy: 68.9 -> 69.47 (+0.57) | Helped |
| Add JointCrop (β=0) | MoCo v3 · ImageNet-1K, 300 epochs | top-1 accuracy: 72.8 -> 73.23 (+0.43) | Helped |
| Add JointCrop (β=0) | SimCLR, BYOL, MoCo v2, and SimSiam · STL-10 | top-1 accuracy: 89.40 -> 90.20 (+0.80); 91.71 -> 92.28 (+0.57); 88.11 -> 89.78 (+1.67); 88.74 -> 89.08 (+0.34) | Helped |
| Add JointCrop (β=0) | SimCLR, BYOL, MoCo v2, and SimSiam · Tiny-ImageNet | top-1 accuracy: 45.25 -> 47.53 (+2.28); 48.91 -> 49.91 (+1.00); 46.07 -> 48.45 (+2.38); 44.17 -> 45.73 (+1.56) | Helped |
| Add JointCrop (β=0) | MoCo v1 and MoCo v2 · ImageNet-100 (COCO pretraining) | accuracy: 59.62 -> 61.92; 56.12 -> 57.92 | Helped |
| Jointly sample brightness parameters with JointColor-b | MoCo v1 · ImageNet-100 | accuracy: 63.18 -> 63.80 | Helped |
| Jointly sample contrast parameters with JointColor-c | MoCo v1 · ImageNet-100 | accuracy: 63.18 -> 63.68 | Helped |
| Set JointCrop β to 2 instead of 0 (ablation) | SimCLR · Tiny-ImageNet | top-1 accuracy: 47.53 -> 46.69 | No clear effect |
| Set JointCrop β to -2 instead of 0 (ablation) | SimCLR · Tiny-ImageNet | top-1 accuracy: 47.53 -> 47.95 | Helped |
| Set JointCrop β to -2 instead of 0 (ablation) | BYOL · Tiny-ImageNet | top-1 accuracy: 49.91 -> 51.02 | Helped |
Surprising findings
- Combining JointCrop and JointBlur can produce overly difficult pairs, so the two are not simply additive and must be applied in sequence or with probability.
- JointBlur gives little effect on MoCo v3 because that method applies Gaussian blur with only 10% probability.
Limitations
- Requires pre-training from scratch, which is computationally expensive, so only limited long-epoch results are shown.
- Applicable only to instance-discrimination contrastive learning, not masked image modeling such as MAE.
- Simultaneous JointCrop and JointBlur can yield overly challenging samples.
Experimental setup
Datasets: ImageNet-1K STL-10 Tiny-ImageNet ImageNet-100 COCO PASCAL VOC
Benchmarks: ImageNet-1K linear evaluation STL-10 Tiny-ImageNet PASCAL VOC detection COCO detection and instance segmentation ImageNet-100
Metrics: top-1 accuracy AP AP50 AP75 training time
Baselines: SimCLR BYOL MoCo v1 MoCo v2 MoCo v3 SimSiam DINO ContrastiveCrop MultiCrop InfoMin
Builds on: SimCLR MoCo BYOL SimSiam ContrastiveCrop
Closest papers by meaning
- A Statistical Theory of Contrastive Learning via Approximate Sufficient Statistics (NeurIPS 2025 · similarity 0.85)
- Diversity Is All You Need for Contrastive Learning: Spectral Bounds on Gradient Magnitudes (NeurIPS 2025 · similarity 0.85)
- On the Similarities of Embeddings in Contrastive Learning (ICML 2025 · similarity 0.85)
- Aligning Visual Contrastive learning models via Preference Optimization (ICLR 2025 · similarity 0.85)
- Variational Supervised Contrastive Learning (NeurIPS 2025 · similarity 0.85)
- Your contrastive learning problem is secretly a distribution alignment problem (NeurIPS 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.