Accelerating constraint-based neural network repairs by example prioritization and selection
摘要
This letter introduces an approach to accelerate constraint-based neural network repairs by example prioritization and selection. The experiments demonstrate the effectiveness of our approach in accelerating constraint-based neural network repairs. Different training methods may lead to changes in the data in Table 1. We repeated the experiment three times, and the data of Table 1 changed very little. In the future, we will explore the effectiveness of the sample selection strategy on other training methods, neural networks, repair approaches, and datasets.