Research Under Review
Rethinking Data Augmentation for Small-Sample Learning: A Stability-Aware Search
Submitted to EUSIPCO 2026. A stability-aware search pipeline for optimizing data augmentation in small-sample learning.
Submission Details
Update (Feb 2026): This paper has been submitted to EUSIPCO 2026 (34th European Signal Processing Conference).
- Paper ID: 1089
- Track 1: VIP - Visual Information Processing (Signal processing for computer vision)
- Track 2: SiG-DML - Signal and Data Analytics for Machine Learning (Neural network learning)
Abstract
In the context of Few-Shot Learning (specifically limited to ~100 samples per class), this research aims to utilize a systematic Pipeline to discover the optimal combination of foundational data augmentation operations. The objective is to verify that a superimposed strategy (multi-op) significantly enhances model accuracy and robustness compared to single-operation baselines.
key Innovations
- Stability-Aware Search: A novel three-stage pipeline initialized based on “Human Priors” restricts the search space to avoid overfitting.
- Multi-Op Superposition: Verifying that rational combinations of operations outperform single operations.
- Joint Search: Moving away from fixed probability constraints to jointly optimize magnitude () and probability ().
Key Dates (EUSIPCO 2026)
- Submission Date: Feb 03, 2026 (Submitted)
- Notification of Acceptance: May 12, 2026 (Expected)
- Camera Ready / Registration: Jun 02, 2026
Methodology
- Phase A: Breadth-first screening to identify effective individual operations.
- Phase B: Depth-first parameter tuning to find the optimal magnitude and probability.
- Phase C: Greedy superposition to construct the final combinatorial strategy.
View the code and detailed experiments on GitHub.