Robert Osazuwa Ness

À propos

I am a Senior Researcher at Microsoft Research AI in the Societal Resilience group. In my research, I explore how to combine causal discovery, causal inference, deep probabilistic modeling, and programming languages to power new capabilities in AI.

I’m particularly interested in building generative models of counterfactual reasoning. For example, humans make heavy use of «would have, could have, should have» counterfactual logic such as «Given that I took the role at MSR and now I’m happy, had I taken the other role, I would have been sad.» This reasoning involves a prediction of an outcome (level of happiness) in an alternate reality (where I took another role) based on outcome data from this reality (I took the MSR role and now I’m happy). A similar pattern in reinforcement learning would be an agent that thinks, «Given that I did this action and got this much reward; how much reward would I have attained had I done a different action?»

I work on machine learning models that can emulate this type of reasoning. One of the challenges is modeling and validation; typically, we can’t get training and test data with hypothetical outcomes in alternate worlds. A second challenge is eliciting and representing causal knowledge about the world in the form of a model; while structural causal models are the theoretical workhorse of counterfactual reasoning, it is challenging to express even the simplest causal systems in this form.

A third problem is how to create primitives for counterfactual reasoning that we can compose with other useful modeling abstractions. For example, when trying to attribute blame for an outcome to an action, one might need to combine a counterfactual inference (how the outcome would have been different given a different action) with some notion of the action’s normality (whether the action was expected or exceptional) and some notion of the outcome’s valency (whether the outcome was good or bad). I believe that in the long run, we need to combine causal primitives with these other abstractions to build AI with higher-order «system 2» cognitive abilities such as common sense, introspection, and theory of mind.

My work targets applications of AI that support human decision-making about complex systems. I am particularly interested in decision-making in economics, energy, and agriculture. My research focuses on developing technology that powers new Azure AI services in these and other verticals.

Before joining MSR, I worked as a research engineer writing production code for probabilistic decision-making under uncertainty. I received my Ph.D. in statistics from Purdue University, where my dissertation focused on Bayesian active learning for causal discovery. I’m a Johns Hopkins SAIS alumni and a graduate of the China campus of the University of Hopkins-Nanjing Center.

To learn more about what I’m working on, read my newsletter.