Research map / Continual learning / Online and replay-based
RCL: Reliable Continual Learning for Unified Failure Detection
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 Reliable Continual Learning (RCL), which first trains a classifier for misclassification detection (CRL/FMFP) and then continually fine-tunes it with LogitNorm under Fisher-weighted regularization and weight-space interpolation to add OOD detection for unified failure detection.
Problem
Unified failure detection, i.e., detecting both misclassified in-distribution examples and OOD examples, where existing OOD detection methods tend to harm misclassification detection and joint training of the two objectives conflicts.
Approach
Decouple the MisD and OOD objectives by learning them in sequence: first equip the model with MisD via CRL or FMFP, then continually fine-tune with LogitNorm while penalizing changes to important parameters (Fisher-weighted EWC-style regularization) and averaging weights along the tuning trajectory.
Key results
- On CIFAR-100 with ResNet110, RCL improves failure detection AUROC over CRL (84.16 to 87.73) and reduces AURC from 139.17 to 123.90.
- RCL on CIFAR-10 WRN-28-10 with FMFP reduces failure detection AURC from 4.75 to 4.62 (approximately; FMFP w/ RCL reports 4.62) and reaches 96.76 AUROC, versus MSP at 93.55 AUROC.
- Improves OOD detection AUROC by 8.18% for CRL and 6.21% for FMFP, per the authors' summary.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply RCL to CRL | CRL with ResNet110 · CIFAR-100 | Failure detection AUROC: 84.16 -> 87.73 | Helped |
| Apply RCL to FMFP | FMFP with ResNet110 · CIFAR-100 | Failure detection AUROC: 84.18 -> 86.75 | Helped |
| Apply RCL to CRL | CRL with WRN-28-10 · CIFAR-100 | Failure detection AUROC: 86.38 -> 88.79 | Helped |
| Apply RCL to FMFP | FMFP with WRN-28-10 · CIFAR-100 | Failure detection AUROC: 88.38 -> 89.34 | Helped |
| Apply RCL to CRL | CRL with ResNet110 · CIFAR-10 | Failure detection AUROC: 93.57 -> 94.82 | Helped |
| Apply RCL to FMFP | FMFP with ResNet110 · CIFAR-10 | Failure detection AUROC: 94.96 -> 95.39 | Helped |
| Apply RCL to CRL | CRL with WRN-28-10 · CIFAR-10 | Failure detection AUROC: 94.68 -> 96.39 | Helped |
| Apply RCL to FMFP | FMFP with WRN-28-10 · CIFAR-10 | Failure detection AUROC: 96.74 -> 96.76 | Helped |
| Add reliable weight consolidation during tuning (ablation) | CRL with ResNet110 · CIFAR-100 | OOD detection AUROC: 71.97 -> 82.24 | Helped |
| Add weight space interpolation (ablation) | CRL with ResNet110 · CIFAR-100 | OOD detection AUROC: 71.97 -> 82.73 | Helped |
| Fine-tune without weight regularization (ablation) | CRL with ResNet110 · CIFAR-100 | OOD detection AUROC: 71.97 -> 84.02 | Helped |
| Use binary interpolation instead of weight space interpolation (ablation) | CRL with ResNet110 · CIFAR-100 | OOD detection AUROC: 71.97 -> 78.90 | Helped |
Surprising findings
- Joint learning of CRL and LogitNorm is worse than the CE baseline for misclassification detection, while sequential learning improves both MisD and OOD detection.
- Fine-tuning only the last block or only BN layers can retain much of the performance, suggesting selective tuning is efficient.
Limitations
- Sequence learning start epoch is hard to choose beforehand.
- Unified failure detection performance remains less than satisfactory relative to ideal, with OOD detection still below LogitNorm alone in sequence learning.
- Evaluated only on CIFAR-scale benchmarks in main text; ImageNet-subset and ViT results are relegated to the appendix.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Textures SVHN Places365 LSUN-C LSUN-R iSUN
Benchmarks: CIFAR-10 CIFAR-100 Textures SVHN Places365 LSUN-C LSUN-R iSUN
Metrics: AURC FPR95 AUROC
Baselines: MSP Energy MaxLogit KNN LogitNorm CRL FMFP
Builds on: CRL LogitNorm elastic weight consolidation FMFP model soups
Closest papers by meaning
- Unsupervised Continual Anomaly Detection with Contrastively-Learned Prompt (AAAI 2024 · similarity 0.80)
- Maintaining Consistent Inter-Class Topology in Continual Test-Time Adaptation (CVPR 2025 · similarity 0.79)
- InCLAD: A Continual Learning Benchmark for Industrial Visual Anomaly Detection (NeurIPS 2026 · similarity 0.79)
- Re-evaluating Continual Learning with Few-Shot Adaptation (NeurIPS 2026 · similarity 0.78)
- H2ST: Hierarchical Two-Sample Tests for Continual Out-of-Distribution Detection (CVPR 2025 · similarity 0.78)
- Provable Contrastive Continual Learning (ICML 2024 · similarity 0.77)
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.