Dr. Valko also teaches the master course Graphs in Machine Learning at l'ENS Paris-Saclay. Dr. Valko is primarily interested in designing algorithms that would require as little human supervision as possible. This means 1) reducing the “intelligence” that humans need to input into the system and 2) minimizing the data that humans need to spend inspecting, classifying, or “tuning” the algorithms. Another important feature of machine learning algorithms should be the ability to adapt to changing environments. That is why Dr. Valko is working in domains that are able to deal with minimal feedback, such as online learning, bandit algorithms, semi-supervised learning, and anomaly detection. Most recently Dr. Valko has worked on sequential algorithms with structured decisions where exploiting the structure leads to provably faster learning.

I will describe adaptive solutions of using graphs for efficiently encoding, discovering, and using the (extra) information that is either explicitly or implicitly present in a given environment. This information can be, smoothness, side observations, state-spaces similarities, or a favorable reward structure which makes the learning faster or easier. I will focus exclusively on online learning and decision making to discuss the necessary tradeoffs that emerge but also when best-of-all-worlds behavior is possible. In particular, I will treat the tradeoffs of representation capacity vs. speed (computational complexity) and capacity vs. learning (statistical complexity) and discuss the optimal allocation of resources. I give specific examples of applying graphs in concrete products (patient data, face recognition, and recommender systems. Finally, I will give solutions for distributed computation with graphs and approximations needed when facing massive data. To sum up, while in the last decade we have been witnessing a huge leap in learning (with) low-level representations such as in vision, the high-level cognition remains a challenge. Graphs offer a natural representation and in this talk, I will attempt to convince you that they can be used to improve systems working with low-level representations.