Explainable AI

The main problem of deep learning is that most applications of this technique are black boxes and hence it is difficult to explain why a decision was taken. This is also true for most ML techniques with varying degree. However, in many applications that affect people, the decision must be explained for legal (e.g., GDPR, the new privacy law of the European Union) and/or ethical reasons (e.g., gender or racial bias). We cover the main challenges and we give examples of current techniques to explain ML models, ending with best practices and an analysis of the impact of AI in the future of humankind.