Plum: Exploration and Prioritization of Model Repair Strategies for Fixing Deep Learning Models

The accuracy of DL models may not meet the user’s expectations. To tackle this problem, existing work proposed diverse approaches, such as using more optimized training processes and training samples to evolve the model structure or parameters of such faulty DL models. In this paper, we present Plum, a novel hyperheuristic approach to fixing deep learning models. Plum generates a set of DL model candidates by applying low-level repair strategies. It then evaluates and prioritizes repair strategies based on their overall fixing effects exhibited by these model candidates and outputs a fixed DL model by applying the top-ranked repair strategy. We also formulate a novel repair strategy to show the compatibility of Plum in incorporating new repair strategies. The experiment on five DL models showed that Plum achieved improvements in test accuracy by 2.49% and 3.11% on the CIFAR-10 and CIFAR-100 datasets over the baselines and outperformed Apricot and MODE, two previous state-of-the-art deep learning repair techniques.