SpaCE-VAE: Sparse and Confident Explanations usingVariational Autoencoders

In the field of explainable AI (XAI), a significant challenge lies in the evaluation of explanation methods. State-ofthe- art techniques identify the importance of input features (e.g., pixels) for the classification, based on untestable assumptions about the model. Lowering the intensity of pixels, identified as irrelevant, typically leads to a change in the predicted class. To address this, we propose SpaCEVAE, a Variational Autoencoder (VAE) designed to generate sparse, confident explanations that are inherently evaluable. SpaCE-VAE produces a sparse representation of an image, putting as many pixels as possible to black, such that the resulting image is still assigned to the same class as the original image (with high confidence).