Research map / Label-efficient and robust learning / Recognition and long-tail
DiffAug: Enhance Unsupervised Contrastive Learning with Domain-Knowledge-Free Diffusion-based Data Augmentation
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 DiffAug, an iterative EM-style framework in which a soft-contrastive semantic encoder and a conditional diffusion generator train each other, so the diffusion model produces positive augmentations without labels or external data.
Problem
Hand-designed augmentations need domain expertise and can distort semantics (e.g., DNA mutations), while generative augmentation methods need labels or large external data, which limits unsupervised contrastive learning in domains such as DNA sequences and bio-feature data.
Approach
A semantic encoder trained with soft contrastive loss supplies condition vectors to a conditional DDPM-style diffusion generator, which then generates positive samples for the encoder's contrastive training; the two are alternately trained (A-step/B-step) so augmentation and representation improve together.
Key results
- On Genomic Benchmarks linear probing, DiffAug reaches 89.1 average accuracy vs. 86.2 for Hyena and 85.4 for SSL+Mixup.
- On bio-feature datasets, DiffAug averages 82.4 vs. 78.5 for DLME+Mixup, with gains of 0.4 to 7.0 points across six datasets.
- On CIFAR-100 linear probing, DiffAug reaches 69.9 vs. 67.4 for DINO (best baseline), and it improves over SimCLR-based baselines on all four vision datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiffAug instead of Hyena | Hyena-tiny DNA representation learning · Genomic Benchmarks (AVE) | linear-probe accuracy: 86.2 -> 89.1 | Helped |
| Use DiffAug instead of DLME+Mixup | Bio-feature representation learning · Bio-feature datasets (AVE) | linear-probe accuracy: 78.5 -> 82.4 | Helped |
| Use DiffAug instead of the strongest listed baseline | Vision representation learning · CF10 | linear-probe accuracy: 92.0 -> 93.4 | Helped |
| Use DiffAug instead of the strongest listed baseline | Vision representation learning · CF100 | linear-probe accuracy: 67.4 -> 69.9 | Helped |
| Use DiffAug instead of the strongest listed baseline | Vision representation learning · STL10 | linear-probe accuracy: 91.8 -> 92.5 | Helped |
| Use DiffAug instead of the strongest listed baseline | Vision representation learning · TinyImageNet | linear-probe accuracy: 47.1 -> 49.7 | Helped |
| Replace the learned semantic condition with supervised one-hot labels (ablation) | DiffAug · CF10 | linear-probe accuracy: 93.4 -> 93.4 | No clear effect |
| Replace the learned semantic condition with random condition vectors (ablation) | DiffAug · CF10 | linear-probe accuracy: 93.4 -> 34.2 | Hurt |
| Remove the diffusion loss (ablation) | DiffAug · CF10 | linear-probe accuracy: 93.4 -> 91.3 | Hurt |
| Replace soft contrastive loss with InfoNCE loss (ablation) | DiffAug · CF10 | linear-probe accuracy: 93.4 -> 92.7 | Hurt |
| Train A-Step and B-Step simultaneously instead of alternating them (ablation) | DiffAug · CF10 | linear-probe accuracy: 93.4 -> 92.9 | Hurt |
Surprising findings
- Naive alternative of using random conditioning vectors instead of encoder-produced ones collapses performance (e.g., CF10 34.2 vs. 93.4).
Limitations
- Training cost is higher due to the diffusion model, and the authors do not scale to larger datasets, only small benchmarks.
- Performance gains on many-class datasets (CF100, TINet) are weaker, with the encoder capturing only some local detail.
- Sensitive to replacement hyperparameter lambda; too large a value causes the encoder to collapse, with 0.1 suggested as default.
Experimental setup
Datasets: Genomic Benchmarks CIFAR-10 CIFAR-100 STL10 TinyImageNet GA1457 SAM561 MC1374 HCL500 PROT579 WARPARIOP
Benchmarks: Genomic Benchmarks CIFAR-10 CIFAR-100 STL10 TinyImageNet GA1457 SAM561 MC1374 HCL500 PROT579 WARPARIOP
Metrics: linear probing accuracy kMeans clustering accuracy
Baselines: DNA-BERT NT Hyena SSL+Translocation SSL+RC SSL+Insertion SSL+Mixup TopoAE PaCMAP hNNE SimCLR+Mixup BYOL+Mixup
Builds on: SimCLR soft contrastive learning DDPM Hyena-DNA DLME
Closest papers by meaning
- DiffAug: A Diffuse-and-Denoise Augmentation for Training Robust Classifiers (NeurIPS 2024 · similarity 0.87)
- A Statistical Theory of Contrastive Learning via Approximate Sufficient Statistics (NeurIPS 2025 · similarity 0.87)
- Do Generated Data Always Help Contrastive Learning? (ICLR 2024 · similarity 0.86)
- BECLR: Batch Enhanced Contrastive Few-Shot Learning (ICLR 2024 · similarity 0.86)
- Learning Diffusion Models with Flexible Representation Guidance (NeurIPS 2025 · similarity 0.86)
- NoiseCLR: A Contrastive Learning Approach for Unsupervised Discovery of Interpretable Directions in Diffusion Models (CVPR 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.