Exploring methods for the generation of visual counterfactuals in the latent space

Abstract

In the field of eXplainable Artificial Intelligence (XAI), the generation of counterfactuals is a promising method for human-interpretable explanations. A counterfactual explanation describes a causal situation in the form: “If X had not occurred, Y would not have occurred”. In this work, we study the generation of visual counterfactuals in the latent space for deep learning image classification models. We explore how to adapt the training environment to facilitate the generation of counterfactuals, combining ideas coming from different fields such as multitasking or generative learning, with the aim of developing more interpretable models. We study well-known counterfactual methods and how to apply them in the latent space. Furthermore, we propose a new way of generating counterfactuals working in the latent space and compare it with the other studied approaches, achieving competitive results.