Curvature Index of Image Samples Used to Evaluate the Interpretability Informativeness

Zhang Zhuo,Xiao Shuai,Xi Meng,Jiabao Wen,Jiachen Yang

DOI: https://doi.org/10.1016/j.engappai.2024.109044

IF: 8

2024-01-01

Engineering Applications of Artificial Intelligence

Abstract:Although there are many thoughts in Artificial Intelligence(AI) interpretability methods, research on the data aspect of artificial intelligence interpretability is still relatively scarce. Previous work has evaluated the value of samples by analyzing the gradient of the sample on the model, but this method does not consider the aftereffect of model optimization, resulting in the evaluation results still having room for improvement. To address this issue, we propose using curvature to evaluate the aftereffect of the sample on model optimization and combining it with the gradient to propose a sample information evaluation method based on the product of gradient and curvature. This method takes into account both the sample’s influence on the model and the sample’s impact on model optimization. Specifically, we introduce a new curvature term in the original information computation formula based on the gradient, which adjusts the amount of information in the sample. The curvature term is calculated as the second derivative of the sample gradient. We conducted performance validation experiments for the algorithm using Active Learning methods on open datasets for image classification tasks and compared our method with several common Active Learning methods. The experimental results show that our method maintains optimal performance in the middle and late stages of Active Learning.