Wilka Carvalho
Reinforcement Learning Researcher & Cognitive Scientist

Predictions of future reward drive many, if not all, of our thoughts and behaviors. But life is filled with countless things, and the events that most reward us are spread few and far between: getting good grades on exams, building meaningful relationships, buying a first home. To work toward these events, we cannot predict reward alone. We must also predict which behaviors and which intermediary events bring the reward closer: that studying a certain way earns good grades, that certain kinds of social interaction build trust and depth, that certain spending habits accumulate into a down payment.

My research aims to characterize the human reinforcement learning algorithm so that we can build machines that help us optimize it. Currently, I am a research fellow in Harvard’s Kempner Institute for the Study of Natural and Artificial Intelligence where I work closely with Sam Gershman. My work spans three themes:

  1. Generalization in naturalistic environments: building deep RL agents that learn efficiently and transfer across tasks in rich, open-ended settings.
  2. Unified models of human and AI learning: computational models that simultaneously predict real human behavior and advance AI performance in naturalistic domains.
  3. AI systems for human flourishing: AI systems that model people—their values, beliefs, goals, and social relationships—to help them live fullfilling lives.

You can find my recent research on my Google Scholar.

I earned my Ph.D. at the University of Michigan, where I studied deep reinforcement learning with Satinder Singh, Honglak Lee, and Richard Lewis, and was supported by the NSF GRFP and a Rackham Merit Fellowship. During my PhD, I was fortunate to spend significant time at DeepMind working with Murray Shanahan, Daniel Zoran, and Danilo Rezende. I got my first stint in ML in a CS M.S. at USC, where I worked with Yan Liu on machine learning for healthcare. Before then, I earned a B.S. in Physics at Stony Brook University where I worked with Axel Drees on computational nuclear physics.

Collaboration and Mentorship

Please feel free to contact me if you’d like to collaborate or be mentored on a research project! While my training is in machine learning, I hope to collaborate broadly with neuroscientists and cognitive scientists. I am also interested in work that makes contact with helping marginalized and minority groups. If you are doing research in this space or organizing any programs and are interested in collaborating, please reach out!

You can reach me at wcarvalho[at]g.harvard.edu

News


Miscellaneous

I also maintain a collection of resources for