DUAL

Dual Alignment Framework for Few-shot Learning with Inter-Set and Intra-Set Shifts

NeurIPS 2025 🏆 Scholar Award

Siyang Jiang, Rui Fang, Hsi-Wen Chen, Wei Ding, Guoliang Xing, Ming-Syan Chen

The Chinese University of Hong Kong (CUHK) • National Taiwan University (NTU)

[Paper] [Code] [Poster] [Results]

Abstract

Few-shot learning (FSL) aims to classify unseen classes with limited labeled samples. However, real-world scenarios often involve distribution shifts between support and query sets, significantly degrading performance. We introduce DUAL, a novel framework that addresses both inter-set shifts (between support and query) and intra-set shifts (within individual sets) through a dual alignment mechanism.

DUAL employs adversarial training to learn robust feature representations and optimal transport to align distributions in embedding space. Our approach includes a pixel-level repairer network that mitigates corruption effects and a dual optimal transport mechanism that performs both intra-set and inter-set alignment. Extensive experiments demonstrate that DUAL achieves state-of-the-art performance across multiple benchmarks under realistic distribution shifts.

Key Contributions

🎯 A New Challenge

Dual Support-Query Shift (DSQS): We formalize the challenging scenario where both support and query sets experience distribution shifts, creating a more realistic but difficult few-shot learning setting.

⚔️ Adversarial Training

Robust Feature Learning: Adversarial training with a generator network creates challenging examples, forcing the encoder to learn more robust and transferable representations.

🔄 Dual Alignment

Two-Stage Optimal Transport: First, intra-set alignment reduces within-set variance. Then, inter-set alignment bridges the gap between support and query distributions using optimal transport.

Experimental Results

Performance on Standard Benchmarks

DUAL achieves state-of-the-art performance across multiple datasets under distribution shift scenarios:

📊 CIFAR-100 Results

Method 1-shot 5-shot
ProtoNet 42.3±0.8 58.1±0.9
PGADA 45.7±0.9 62.4±0.8
DUAL (Ours) 48.9±0.8 65.7±0.7

📊 mini-ImageNet Results

Method 1-shot 5-shot
ProtoNet 49.4±0.8 68.2±0.7
PGADA 52.3±0.9 71.5±0.6
DUAL (Ours) 55.8±0.7 74.2±0.6

Key Findings

🎯 Robustness to Shifts

DUAL maintains superior performance even under severe distribution shifts, demonstrating the effectiveness of our dual alignment approach.

⚡ Computational Efficiency

Despite the sophisticated alignment mechanism, DUAL maintains reasonable computational overhead with significant accuracy improvements.

🔧 Ablation Studies

Each component (repairer, adversarial training, dual OT) contributes meaningfully to the overall performance, validating our design choices.

Citation

If you use DUAL in your research, please cite our paper:

@inproceedings{jiang2025dual,
  title     = {Dual Alignment Framework for Few-shot Learning with Inter-Set and Intra-Set Shifts},
  author    = {Jiang, Siyang and Fang, Rui and Chen, Hsi-Wen and Ding, Wei and Xing, Guoliang and Chen, Ming-Syan},
  booktitle = {Proceedings of the 39th Annual Conference on Neural Information Processing Systems (NeurIPS 2025)},
  year      = {2025}
}

Related Publications

Contact Information

For questions or collaborations: