Submission Outcome

Update (May 2026): This paper was not accepted by EUSIPCO 2026. The result was disappointing, but the submission exposed areas where the methodology and experimental validation can be strengthened. I plan to carry those lessons into future research.

• 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 ($m$) and probability ($p$).

Key Dates (EUSIPCO 2026)

• Submission Date: Feb 03, 2026 (Submitted)
• Decision: May 12, 2026 (Not accepted)
• Camera Ready / Registration: Not applicable

Methodology

1. Phase A: Breadth-first screening to identify effective individual operations.
2. Phase B: Depth-first parameter tuning to find the optimal magnitude and probability.
3. Phase C: Greedy superposition to construct the final combinatorial strategy.

View the code and detailed experiments on GitHub.