Positive-Congruent Training: Towards Regression-Free Model Updates
This work is significant for practitioners deploying updated AI models, as it provides a method to reduce undesirable behavioral inconsistencies (negative flips) for end-users.
The paper addresses the problem of "negative flips" in image classification, where a new model incorrectly classifies samples that an old model classified correctly. They propose Focal Distillation, which reduces negative flips while maintaining accuracy by weighting correctly classified samples more heavily. Using an ensemble of reference models further reduces negative flips without impacting accuracy.
Reducing inconsistencies in the behavior of different versions of an AI system can be as important in practice as reducing its overall error. In image classification, sample-wise inconsistencies appear as "negative flips": A new model incorrectly predicts the output for a test sample that was correctly classified by the old (reference) model. Positive-congruent (PC) training aims at reducing error rate while at the same time reducing negative flips, thus maximizing congruency with the reference model only on positive predictions, unlike model distillation. We propose a simple approach for PC training, Focal Distillation, which enforces congruence with the reference model by giving more weights to samples that were correctly classified. We also found that, if the reference model itself can be chosen as an ensemble of multiple deep neural networks, negative flips can be further reduced without affecting the new model's accuracy.