Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Meta Evidential Transformer for Few-Shot Open-Set Recognition
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 Meta Evidential Transformer (MET), which combines evidential open-set loss with opponent closed-set classes, an evidence-to-variance ratio, and evidence-guided cross-attention to improve few-shot open-set recognition.
Problem
Few-shot open-set recognition: detecting unseen-class instances using few labeled closed-set examples, especially hard open-set samples that resemble closed-set classes.
Approach
Uses semantically similar closed-set classes as training-time opponent open-set classes with an evidential KL loss pushing their evidence toward zero, plus an evidence-to-variance ratio at inference to detect hard tasks and reweight transformer attention (evidence-guided cross-attention) for better separation.
Key results
- On MiniImageNet 5-way 1-shot, MET reaches 76.93% AUROC vs 67.37% for SnaTCHer; 5-shot 84.90% vs 77.99%.
- On TieredImageNet 5-way 1-shot, MET reaches 78.77% AUROC vs 71.00% for SnaTCHer.
- Ablation on MiniImageNet 1-shot: transformer alone 67.37, +evidential loss 74.35, +EVR 76.93 AUROC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MET instead of Glocal | MiniImageNet 5-way 1-shot · MiniImageNet 5-way 1-shot | AUROC: 73.41 ± 0.83 -> 76.93 ± 0.59 | Helped |
| Use MET instead of Glocal | MiniImageNet 5-way 5-shot · MiniImageNet 5-way 5-shot | AUROC: 83.45 ± 0.57 -> 84.90 ± 0.41 | Helped |
| Use MET instead of TANE | TieredImageNet 5-way 1-shot · TieredImageNet 5-way 1-shot | AUROC: 74.89 ± 0.64 -> 78.77 ± 0.46 | Helped |
| Use MET instead of TANE | TieredImageNet 5-way 5-shot · TieredImageNet 5-way 5-shot | AUROC: 80.45 ± 0.49 -> 84.37 ± 0.35 | Helped |
| Use MET instead of Glocal | Cifar100 5-way 1-shot · Cifar100 5-way 1-shot | AUROC: 57.71 ± 0.79 -> 61.76 ± 0.60 | Helped |
| Use MET instead of TANE | Cifar100 5-way 5-shot · Cifar100 5-way 5-shot | AUROC: 63.08 ± 0.58 -> 66.17 ± 0.54 | Helped |
| Use MET instead of SnaTCHer | Caltech101 5-way 1-shot · Caltech101 5-way 1-shot | AUROC: 62.37 ± 0.63 -> 64.85 ± 0.55 | Helped |
| Use MET instead of SnaTCHer | Caltech101 5-way 5-shot · Caltech101 5-way 5-shot | AUROC: 67.35 ± 0.53 -> 72.12 ± 0.47 | Helped |
| Add the evidential open-set loss (ablation) | MET · MiniImageNet 5-way 1-shot | AUROC: 67.37 -> 74.35 | Helped |
| Add the evidential open-set loss (ablation) | MET · MiniImageNet 5-way 5-shot | AUROC: 77.99 -> 81.47 | Helped |
| Add EVR augmented inference (ablation) | MET with evidential loss · MiniImageNet 5-way 1-shot | AUROC: 74.35 -> 76.93 | Helped |
| Add EVR augmented inference (ablation) | MET with evidential loss · MiniImageNet 5-way 5-shot | AUROC: 81.47 -> 84.90 | Helped |
Surprising findings
- Original data split results show MET only marginally above SnaTCHer on 1-shot (73.20 vs 72.98), suggesting gains depend on the altered split with similar open/closed classes.
Limitations
- Relies on the existence of semantically similar closed-set classes to serve as opponent classes; requires semantic analysis or a pretrained classifier to select them.
- Evaluation uses a modified data split where open and closed classes are chosen to be similar, departing from standard splits; gains on original split are small.
- Method relies on an EVR threshold chosen to exceed all task EVR values, and hyperparameters (lambda, epsilon) are tuned via sensitivity analysis.
Experimental setup
Datasets: MiniImageNet TieredImageNet CIFAR100 Caltech101
Benchmarks: MiniImageNet TieredImageNet CIFAR100 Caltech101
Metrics: AUROC closed-set accuracy
Baselines: ProtoNet RelationNet OpenMax FEAT PEELER SnaTCHer TANE Glocal OSLO
Builds on: FEAT SnaTCHer evidential deep learning PEELER ProtoNet
Closest papers by meaning
- Few-Shot Open-Set Recognition by Transformation Consistency (CVPR 2021 · similarity 0.91)
- Few-Shot Open-Set Recognition Using Meta-Learning (CVPR 2020 · similarity 0.89)
- Towards Evidential and Class Separable Open Set Object Detection (AAAI 2024 · similarity 0.87)
- One Meta-tuned Transformer is What You Need for Few-shot Learning (ICML 2024 · similarity 0.85)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.84)
- Fine-Grained Prototypes Distillation for Few-Shot Object Detection (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.