Visualizing the best Noise for Classification

Recent studies have shown that adding noise to sample points can actually benefit classifiers,which raises an interesting question: what does the best noise actually look like? To address this question, we designed a simple objective function aimed at applying the minimal noise to a sample so that its loss is reduced as much as possible.By optimizing this objective, we can find the best noise for the sample and visualize it, while also comparing the accuracy improvement after adding this noise. However, this noise generation method relies on the true labels, which is impractical in real-world tasks. Therefore, we aim to construct a dataset of image and best noise pairs to predict noise from images, which will be our future work.