Improving Model Safety by Targeted Error Correction
For developers of safety-critical ML systems, this provides a post-hoc method to improve model safety without retraining, though the gains are incremental over existing confidence-based approaches.
The paper introduces a dual-classifier GBDT pipeline to correct high-risk misclassifications in safety-critical applications, achieving a 34.1% reduction in dangerous errors for skin lesion diagnosis and 12.57% for prostate histopathology, with negligible inference latency overhead (under 2%).
The widespread adoption of machine learning in critical applications demands techniques to mitigate high-consequence errors. Our method utilizes a dual-classifier GBDT pipeline to distinguish routine human-like errors from high-risk non-human misclassifications. Evaluated across three domains, animal breed classification, skin lesion diagnosis (ISIC 2018), and prostate histopathology (SICAPv2), our framework demonstrates robust safety improvements. To address real-world deployment concerns, our results confirm the pipeline introduces negligible inference latency (1.60% overhead for the animal dataset, 1.84% for ISIC, and 1.70% for SICAPv2) while outperforming traditional Maximum Class Probability (MCP) baselines in correction precision. Our conservative correction strategy successfully reduced dangerous non-human errors by 34.1% in ISIC and 12.57% in SICAPv2, improving super-class diagnostic safety to 90.41% and 92.13% respectively. This proves that safety-critical reliability can be substantially enhanced post-hoc without expensive model retraining. keywords: Error Analysis, Post-hoc Correction, Trustworthy AI.