ReliableNet: A Chance-Constrained Approach to Trustworthy Classification in Deep Learning
The paper targets the specific failure where a classifier is confident at the same time it is wrong.
ReliableNet trains with a user-set budget for that joint confident-wrong probability, instead of only optimizing average loss. The authors say their smooth conservative approximation preserves the original constraint at the population level. In tests across tabular and image datasets, they report ReliableNet as the only compared method certified within budget for every in-distribution dataset and seed. They also report lower empirical joint confident-wrong rates under several dataset shifts while staying competitive on accuracy, coverage, calibration, and selective prediction. ArXiv · AI/CL/LG's note
ReliableNet trains with a user-set budget for that joint confident-wrong probability, instead of only optimizing average loss. The authors say their smooth conservative approximation preserves the original constraint at the population level. In tests across tabular and image datasets, they report ReliableNet as the only compared method certified within budget for every in-distribution dataset and seed. They also report lower empirical joint confident-wrong rates under several dataset shifts while staying competitive on accuracy, coverage, calibration, and selective prediction. ArXiv · AI/CL/LG's note
score 4