Reducing overinterpretation in deep neural networks through SIS-dependent Label Smoothing

In image classification, overinterpretation occurs when a classifier finds strong class-evidence in regions of an image that contain no semantically meaningful features. This issue can put into question the trustworthiness and reliability of image classification models. To diagnose the presence of overinterpretation, Sufficient Input Subset (SIS) procedures are used. These procedures can identify subsets of pixels within an image that are sufficient to maintain the model's initial prediction, and which constitute valid statistical signals that the model relies on. We focus on conducting a proper analysis of the SIS obtained through various SIS procedures to gain insight into the model's behavior. Moreover, we introduce the SIS-dependent Label Smoothing method to address the issue of overinterpretation. This procedure involves penalizing the model whenever an SIS procedure reveals patterns that are indicative of overinterpretation, and achieves this by applying label smoothing on a per-image basis, taking into account the characteristics of the SIS generated specifically for that image. We find this method effective for reducing overinterpretation, as it becomes necessary to have more unmasked pixels in the SIS to maintain the prediction made by the model. Furthermore, this reduction in overinterpretation isn't limited to the patterns found through the particular SIS procedure employed during training but applies to pixel-subsets generated from any SIS procedure.