Associate Professor of Computer Science
Joined department: Fall 2015
Ribeiro's research focuses on endowing machine learning algorithms with the ability to learn robust invariant representations for relational and temporal data for both associational and causal tasks.
Invariances are one of the most consequential forms of prior knowledge in mathematics, physics, statistics, and also machine learning. An object is invariant if it is the same under some pre-specified set of transformations.
For instance, graphs and tensors often are the same under index permutations (isomorphism), while temporal and spatial invariances often imply stationarity, and robustness to adversarial attacks is an invariance to adversarial perturbations.
Causal relationships between variables can be described as the absence of an invariance in their relationship. Whether designing recommender systems for social networks, building robots that can reason about the relationships between objects in their surrounding environment, learning to generate new drugs, learning causal relationships, or extracting logical rules from data examples, invariances are essential for robust automated learning.
The outcome of Ribeiro's lab research has had far-reaching implications, from a principled framework to perform counterfactual tasks (predicting what-if scenarios), to endowing machine learning models with world-knowledge.