Research map / Label-efficient and robust learning / Recognition and long-tail
SAFLEX: Self-Adaptive Augmentation via Feature Label Extrapolation
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 SAflex, a bilevel-motivated, greedy online method that learns per-sample weights and soft labels for augmented samples from any upstream augmentation pipeline, solved via a linear program with a gradient-matching objective.
Problem
Augmented samples from upstream pipelines (hand-crafted, learned, or diffusion-generated) can carry feature-distribution noise and label errors that mislead training, and existing learnable augmentation methods are costly or restricted to differentiable transformations.
Approach
Instead of modifying augmented features, learn only low-dimensional sample weights in [0,1] and soft labels in the simplex for each augmented sample, chosen greedily per iteration by maximizing the inner product between the augmented-batch gradient and a validation gradient; the resulting linear program has closed-form solutions (one-hot label, weight 0/1), plus a label-retention penalty and Gumbe
Key results
- On MedMNIST, SAflex improves RandAugment and Mixup with gains up to about 3.6% (e.g., OrganA 94.18 to 95.32 for RandAug).
- On seven tabular datasets, SAflex with CutMix improves MLP accuracy on Credit from 73.68 to 74.61 and on Volkert from 61.20 to 61.51, with consistent gains across datasets.
- On diffusion-generated augmentations, SAflex improves Img2Img (w/ finetune) fine-grained accuracy from 72.01 to 73.61 and OOD accuracy from 70.49 to 72.83.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply SAFlex to RandAugment | ResNet-18 on MedMNIST · PathMNIST, DermaMNIST, TissueMNIST, BloodMNIST, OCTMNIST, OrganAMNIST, OrganCMNIST, OrganSMNIST | accuracy | Mixed |
| Apply SAFlex to CutMix | MLP tabular classifiers · Appetency, Arrhythmia, Click, Credit, QASR, Shrutime, Volkert | score | Helped |
| Apply SAFlex to CutMix | SAINT tabular classifiers · Appetency, Arrhythmia, Click, Credit, QASR, Shrutime, Volkert | score | Mixed |
| Apply SAFlex to InstructPix2Pix augmentations | ResNet-50 diffusion augmentation · Fine-grained classification and OOD generalization | accuracy | Helped |
| Apply SAFlex to Img2Img augmentations without fine-tuning | ResNet-50 diffusion augmentation · Fine-grained classification and OOD generalization | accuracy | Helped |
| Apply SAFlex to Img2Img augmentations with fine-tuning | ResNet-50 diffusion augmentation · Fine-grained classification and OOD generalization | accuracy | Helped |
| Apply SAFlex to RandAugment | CLIP fine-tuned with FLYP · iWildCam ID and OOD, with and without ensembling | accuracy | Helped |
| Use Crop augmentation instead of no augmentation | ResNet-18 on MedMNIST · MedMNIST | accuracy | Hurt |
| Use Soft-Augmentation with Crop instead of Crop alone | ResNet-18 on MedMNIST · MedMNIST | accuracy | Mixed |
Surprising findings
- Soft-Augmentation with crop augmentation underperforms the no-augmentation baseline on most MedMNIST datasets.
- The derived closed-form solutions yield only hard (0/1) weights and one-hot labels, which may limit generalization in under-parameterized regimes, motivating the Gumbel-softmax and label-retention modifications.
Limitations
- Requires a substantial, high-quality validation set; a suboptimal set limits effectiveness.
- Performance depends on the choice of upstream augmentation method.
- Incurs computational overhead from frequent gradient evaluations (about 42% longer epochs on MedMNIST).
Experimental setup
Datasets: MedMNIST (8 datasets) tabular datasets (Appetency, Arrhythmia, Click, Credit, QSAR, Shrutime, Volkert) CUB subset iWildCam (WILDS)
Benchmarks: MedMNIST tabular benchmarks (Appetency, Arrhythmia, Click, Credit, QSAR, Shrutime, Volkert) CUB fine-grained subset iWildCam OOD subset
Metrics: test accuracy
Baselines: No Aug RandAugment Mixup CutMix Soft-Augmentation LP-A3 Text2Img LP-FT FLYP+RandAugment
Builds on: AutoAugment RandAugment Mixup CutMix Soft-Augmentation FLYP
Closest papers by meaning
- Provable Benefit of Cutout and CutMix for Feature Learning (NeurIPS 2024 · similarity 0.85)
- DiffuseMix: Label-Preserving Data Augmentation with Diffusion Models (CVPR 2024 · similarity 0.85)
- Data Augmentation with Diffusion for Open-Set Semi-Supervised Learning (NeurIPS 2024 · similarity 0.85)
- UtilGen: Utility-Centric Generative Data Augmentation with Dual-Level Task Adaptation (NeurIPS 2025 · similarity 0.84)
- Effective Data Augmentation With Diffusion Models (ICLR 2024 · similarity 0.84)
- An Augmentation-Aware Theory for Self-Supervised Contrastive Learning (ICML 2025 · similarity 0.84)
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.