Finding NEM-U: Explaining unsupervised representation learning through neural network generated explanation masks

Figure 1. NEM-U explaining how important various parts of an image are for feature extractors trained using different methods. The feature extractor trained using supervised learning focuses on the ears and face of the object. The DINO pretrained feature extractor considers the entirety of all animals. The SwAV and SimCLR feature extractors look at both the focused object and the background, where SwAV is more focused than SimCLR. Explanations are generated without optimization on the image (taken from VOC data set).