Interpretable and Explainable Machine Learning - from Prediction to Knowledge

Undergraduate / Graduate

ML is already ubiquitous, revolutionizing our lives in all domains. Precisely because of this, its new frontier is occupied with finding ways to ensure that we truly understand and exert control over how our techniques predict what they predict, and generate the results that they generate. This movement beyond being satisfied with just the accuracy of results is critical for a couple of reasons. First, as the wise say, with great power comes great responsibility - since our revolutionary ML techniques only learn from what we feed them, they reproduce (and even exacerbate) biases present in our data. To counter this, we need to be able to interpret and then intervene in how Machine Learning learners are learning what they learn. Second, by being able to interpret, explain and modify how algorithms learn what they learn from our datasets (not merely how algorithms mechanically work), we can uncover relevant latent knowledge hidden in such data. In this course, we start by discussing what Interpretable or Explainable ML even means to different audiences. Next, students will be introduced to several types of state-of-the-art techniques that have been proposed to increase the interpretability/explainability of ML models. Along the way, students will be presented to the discussion on how interpretable or explainable models can help fight biases, improve the fairness, trustworthiness or reliability of predictions, and ensure the ethical use of ML as it continues to change our lives.