Local Path Integration for Attribution

Path attribution methods are a popular tool to interpret visual model's prediction on an input. They integrate model gradients for the input features over path defined between and reference, thereby satisfying certain desirable theoretical properties. However, their reliability hinges choice of reference. Moreover, they do not exhibit weak dependence input, which leads counter-intuitive feature mapping. We show that path-based can account property by choosing reference from local distribution devise method identify propose technique stochastically paths references sampled distribution. Our integration (LPI) is found consistently outperform existing techniques when evaluated deep models. Contributing ongoing search reliable evaluation metrics interpretation methods, we also introduce DiffID metric uses relative difference insertion deletion games alleviate shift problem faced metrics. code available at https://github.com/ypeiyu/LPI.