Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

FSL-Rectifier: Rectify Outliers in Few-Shot Learning via Test-Time Augmentation

AAAI 2025 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Alignment-based few-shot metric learning · 0 citations (OpenAlex)

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 FSL-Rectifier, a training-free test-time augmentation that generates extra test-class images via a GAN-based image combiner, picks neighbours with a learned selector, and averages embeddings to correct outlier queries/supports for trained FSL models.

Problem

Few-shot learning models suffer generalization errors when outlier query or support images appear at inference time, and existing augmentation mostly happens at training time.

Approach

Test-time augmentation: combine each test image's style with the shape of a suitable train-class neighbour via a pretrained generative image combiner, then average embeddings of the original and generated copies so predictions rely on more typical representations; a neighbour selector ranks candidates without generating images.

Key results

• On Animals with Euclidean ProtoNet, FSL-Rectifier raises 5-way-1-shot accuracy from 54.64% to 57.90%; on FEAT from 46.86% to 53.28%.
• Inverse neighbour selector drops ProtoNet accuracy to 51.08%, and random neighbour selection gives 56.70%, showing the selector matters.
• Average maximum pairwise feature-difference norm decreases from 21.37 to 19.08 after rectification, consistent with the theory.

Measured effects reported in the paper

Surprising findings

• Naive Mix-Up and standard augmentations (Rotate, Affine, ColorJitter) often fail to improve or degrade the baseline, while Oracle with real extra samples reaches about 80%.
• Support-only augmentation can hurt performance (ProtoNet drops from 54.64% to 53.46% with 1 query copy and no support copy), so support and query augmentation counts must match.

Limitations

• Image generation quality is limited, and poor generations can occur, requiring the neighbour selector.
• Gains are modest on some models (e.g., Mammals with DeepSets and Mix-Up/affine show no improvement), and the theory assumes an SVM classifier on pairwise feature differences rather than the actual FSL classifier.
• Augmentation counts must be matched between support and query, and the method adds inference-time generation cost.

Experimental setup

Datasets: Animals (from ImageNet) Mammals

Benchmarks: Animals Mammals

Metrics: 5-way-1-shot accuracy 95% confidence interval

Baselines: No Augmentation Oracle Rotate ColorJitter Affine Mix-Up

Builds on: ProtoNet FEAT DeepSets CycleGAN MAML

Closest papers by meaning

• Provably Improving Generalization of Few-shot models with Synthetic Data (ICML 2025 · similarity 0.86)
• Few-Shot Learner Generalizes Across AI-Generated Image Detection (ICML 2025 · similarity 0.85)
• Few-shot Learner Parameterization by Diffusion Time-steps (CVPR 2024 · similarity 0.85)
• BECLR: Batch Enhanced Contrastive Few-Shot Learning (ICLR 2024 · similarity 0.85)
• Data Distribution Distilled Generative Model for Generalized Zero-Shot Recognition (AAAI 2024 · similarity 0.85)
• Strong Baselines for Parameter-Efficient Few-Shot Fine-Tuning (AAAI 2024 · similarity 0.85)