Abstract
Artificial Intelligence is being applied in a multitude of scenarios that are sensitive to the human user, i.e., medical diagnosis, granting loans, human resources management, among many others. Behind most of these Artificial Intelligence tools is a pattern recognition model generated by Machine Learning. To do this, it is necessary to start from a dataset that characterizes the problem under study, and “train” this model to represent the former information through different mathematical approximations. Thus, when sensitive applications and mathematical models are placed in the same equation, mistrust arises about the correct functioning of Artificial Intelligence systems. This is the main reason behind which the model makes one decision and not another. The answer lies in the interpretability or transparency of the model itself, i.e., that its components are directly understandable by the human user. When this is not possible, a posteriori explainability mechanisms are used to facilitate knowledge of which variables or characteristics the model has considered. Throughout this chapter, we will introduce the current trends to achieve trustworthy Artificial Intelligence. We will expose the components that allow a model to be transparent, as well as the existing techniques to explain more complex models such as those based on Deep Learning. Finally, we will expose some prospects that can be considered to keep improving the explanations and to allow a wider use of Machine Learning solutions in all fields of application.