Research map / Label-efficient and robust learning / Recognition and long-tail
Exploring Diverse Representations for 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 MEDAF, a discriminative multi-expert network whose experts are pushed to learn mutually different CAM attention maps and whose logits are adaptively fused for open set recognition.
Problem
Open set recognition requires classifying known classes while rejecting unknown samples; generative OSR methods perform well but are computationally expensive and unstable on complex or large-scale tasks.
Approach
Theoretically argues that supplementary class-relevant representations outside the focal region reduce open space risk, and realizes them discriminatively via multiple experts sharing shallow layers, with an attention diversity regularizer on CAMs and a gating network that fuses expert logits; rejection uses a combined logit and feature-norm score.
Key results
- On unknown detection, MEDAF reaches AUROC 0.800 on Tiny-ImageNet, the best among compared methods (DIAS 0.731), and 0.960 on CIFAR+10.
- On ImageNet-1k (100 known classes), MEDAF achieves AUROC 0.928 versus 0.836 for ARPL.
- On CIFAR10 OOD detection with SVHN as out-of-distribution, MEDAF gets AUROC 0.991 versus 0.966 for ARPL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MEDAF instead of DIAS for unknown detection | ResNet18 · Tiny-ImageNet unknown detection | AUROC: 0.731 -> 0.800 | Helped |
| add attention diversity loss to the single-expert baseline (ablation) | MEDAF · CIFAR10 known / ImageNet-R unknown | macro-F1: 0.944 -> 0.977 | Helped |
| add attention diversity loss to the single-expert baseline (ablation) | MEDAF · CIFAR10 known / ImageNet-F unknown | macro-F1: 0.915 -> 0.951 | Helped |
| add attention diversity loss to the single-expert baseline (ablation) | MEDAF · CIFAR10 known / LSUN-R unknown | macro-F1: 0.952 -> 0.987 | Helped |
| add attention diversity loss to the single-expert baseline (ablation) | MEDAF · CIFAR10 known / LSUN-F unknown | macro-F1: 0.901 -> 0.946 | Helped |
| replace average expert prediction with gating-network weights (ablation) | MEDAF with diversity loss · CIFAR10 known / ImageNet-R unknown | macro-F1: 0.977 -> 0.983 | Helped |
| replace average expert prediction with gating-network weights (ablation) | MEDAF with diversity loss · CIFAR10 known / ImageNet-F unknown | macro-F1: 0.951 -> 0.953 | Helped |
| replace average expert prediction with gating-network weights (ablation) | MEDAF with diversity loss · CIFAR10 known / LSUN-R unknown | macro-F1: 0.987 -> 0.990 | Helped |
| replace average expert prediction with gating-network weights (ablation) | MEDAF with diversity loss · CIFAR10 known / LSUN-F unknown | macro-F1: 0.946 -> 0.950 | Helped |
| increase the number of experts (ablation) | MEDAF · multiple datasets | closed-set accuracy and AUROC | Mixed |
Surprising findings
- Generative OSR methods were found computationally infeasible to train on ImageNet in prior work, while the discriminative MEDAF scales with little added cost.
- Adding the attention diversity loss alone raises CIFAR10 open-set results notably over a single expert baseline (e.g., LSUN-R 0.952 to 0.987).
Limitations
- Number of experts shows diminishing returns beyond a few experts.
- Expert number and block configuration must be tuned to backbone and demand.
Experimental setup
Datasets: CIFAR10 CIFAR100 SVHN Tiny-ImageNet LSUN ImageNet
Benchmarks: CIFAR10 CIFAR+10 CIFAR+50 SVHN Tiny-ImageNet ImageNet-1k
Metrics: AUROC macro-F1 DTACC AUIN AUOUT TNR@TPR95 closed-set accuracy GFLOPs
Baselines: Softmax C2AE CGDL RPL PROSER CVAECap ARPL NAS-OSR DIAS CSI GCPL ConOSR
Builds on: mixture of experts class activation mapping ARPL ConOSR
Closest papers by meaning
- Frequency Shuffling and Enhancement for Open Set Recognition (AAAI 2024 · similarity 0.88)
- From Coarse to Fine-Grained Open-Set Recognition (CVPR 2024 · similarity 0.88)
- All Beings Are Equal in Open Set Recognition (AAAI 2024 · similarity 0.86)
- Improving Open Set Recognition via Visual Prompts Distilled from Common-Sense Knowledge (AAAI 2024 · similarity 0.85)
- Overcoming the Pitfalls of Vision-Language Model Finetuning for OOD Generalization (ICLR 2024 · similarity 0.84)
- OpenViewer: Openness-Aware Multi-View Learning (AAAI 2025 · 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.