1S-DAug: One-Shot Data Augmentation for Robust Few-Shot Generalization

2025-12-04 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, extended

Summary

1S-DAug is a novel one-shot generative data augmentation operator designed to enhance few-shot learning (FSL) model generalization to novel classes, particularly at test time. Unlike traditional test-time augmentations that often lack effectiveness, 1S-DAug synthesizes diverse yet faithful image variants from a single example. This is achieved by combining geometric perturbations, controlled noise injection, and a denoising diffusion process conditioned on the original image. The generated images are then encoded and aggregated with the original image into a combined representation, leading to more robust FSL predictions. Implemented as a training-free, model-agnostic plugin, 1S-DAug consistently improves FSL performance across four different datasets, including a proportional accuracy improvement exceeding 10% on the miniImagenet 5-way-1-shot benchmark, without requiring any model parameter updates.

Key takeaway

For AI Engineers and Research Scientists working on few-shot learning, 1S-DAug offers a significant advantage by improving model generalization to unseen classes without requiring model retraining or parameter updates. You should consider integrating this training-free, model-agnostic plugin into your FSL pipelines, especially for applications like medical imaging or autonomous driving where data scarcity is common. This approach can yield substantial accuracy gains, as demonstrated by the >10% improvement on miniImagenet, by enhancing the diversity and robustness of test-time representations.

Key insights

1S-DAug improves few-shot learning generalization by generating diverse, faithful image variants at test time via diffusion.

Principles

• Diversity and accuracy are crucial for effective data augmentation.
• Test-time augmentation directly addresses novel class generalization.
• Feature aggregation enhances robustness for non-parametric classifiers.

Method

1S-DAug applies geometric transformations, injects controlled noise, and then denoises via an attention-guided diffusion process conditioned on the original image to synthesize variants for aggregation.

In practice

• Use 1S-DAug as a training-free plugin for FSL models.
• Adjust noise level to balance diversity and faithfulness.
• Aggregate features from original and augmented views for prediction.

Topics

• Few-Shot Learning
• Data Augmentation
• Diffusion Models
• Generative Augmentation
• Test-Time Augmentation

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• ZeroDiff++: Substantial Unseen Visual-semantic Correlation in Zero-shot Learning — ZeroDiff++ enhances zero-shot learning by mitigating spurious visual-semantic correlations through diffusion-based generation and adaptation.
• Data augmented bootstrap: Unifying confidence interval construction by approximate invariance — Data Augmented Bootstrap (DAB) unifies confidence interval construction by leveraging approximate data invariances, extending classical and modern methods.
• Can You Build Effective AI Models with Small Datasets? — Effective AI models can be built with small datasets by prioritizing data quality and employing specific technical strategies.
• EAGT: Echocardiography Augmentation for Generalisability and Transferability — Geometric data augmentations significantly improve echocardiography model generalizability across diverse datasets, outperforming intensity-based methods.
• Accelerating Disaggregated RL for Visual Generative LLMs with Diffusion-Based Parallelism and Trainer-Assisted Generation — DigenRL disaggregates RL for visual generative LLMs, using diffusion-based parallelism and trainer assistance for 1.56-2.10x throughput.
• A Data-Augmented Contrastive Learning Approach to Nonparametric Density Estimation — DANNCE offers an efficient, consistent, and automatically normalized nonparametric density estimation via contrastive learning and data augmentation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.