From explainable machine learning to causality
Anisio Lacerda – Universidade Federal de Minas Gerais, Brasil
Tracks de Socioeconomía Computacional y de Inteligencia Artificial
Nota: El curso será dictado en Inglés
Machine learning has been successfully adopted in many areas, such as medicine, e-commerce, and education. For instance, recent advances of diabetic retinopathy diagnosis has reached a precision 0.991 (in terms of AUC, 1.0 is best). In e-commerce, especially in recommendation systems, technology firms, such as Amazon, Google, and Netflix, are increasing profits by providing better interaction among their services and users. Finally, machine learning has also been used in creative tasks, such as supporting the creation of new signs for deaf students. These are few examples of the usage of machine learning in prediction tasks. However, the most successful machine learning techniques, e.g. deep learning, are considered to be “black-boxes” given that it is difficult to understand the reasons behind given predictions. A key challenge of statistical learning models is the lack of an explicit knowledge representation. In other words, machine learning, as an isolated prediction task, is unable to explain the relationship between cause and effects. In this presentation, we aim to review basic concepts of machine learning, walk through explainable machine learning techniques and show how causal inference may help to understand machine learning predictions.
• Machine learning review and applications: regression, classification, clustering, multi-armed bandits, recommender systems, computational advertising • Interpretable machine learning: motivation, taxonomy, model and outcome explanations • Causality: motivation, probabilistic programming, DAGs, interventions, do-calculus, counterfactuals, deep causal latent variable models.
Basic knowledge on probability (e.g., random variables, joint probability distributions, conditional probabilities distributions, Bayes rule).