One of the defining characteristics of human intelligence is the ability to solve a wide range of problems. Current AI and machine learning techniques, by contrast, excel only when used to tackle narrowly defined tasks. In this talk, I will discuss novel learning algorithms that can solve large sets of diverse real-life problems that are not known in advance, while ensuring that the learning process is safe. I will focus on two key challenges: (1) how to construct general-purpose reinforcement learning algorithms capable of autonomously decomposing complex tasks into simpler sub-problems, for which specialized reusable and composable skills be can be learned; and (2) how to ensure that these skills are learned in a way that meets user-specified safety requirements with high probability. These are fundamental questions that underlie the gap between what artificial intelligence agents can--in principle--do and what we can effectively get them to do, given our current algorithms. Finally, I will discuss future research directions towards enabling algorithms to solve real-world tasks, in the home and in the workplace, in a safe way, and with as little human intervention as possible.
Bruno C. da Silva is an assistant professor at the Computer Science Department of the Federal University of Rio Grande do Sul, in Brazil. Prior to that, he was a postdoctoral associate at the Aerospace Controls Laboratory at MIT. He received his Ph.D. in Computer Science from the University of Massachusetts, under the supervision of Prof. Andrew Barto. Both his MSc. and B.S. cum laude degrees are in Computer Science from the Federal University of Rio Grande do Sul. Bruno has worked, on several occasions, as a visiting researcher at the Laboratory of Computational Neuroscience, in Rome, Italy, developing novel control algorithms for humanoid robots. He has also worked at Adobe Research, in California, developing large-scale machine learning techniques for digital marketing optimization. Bruno's research interests lie in the intersection of machine learning, reinforcement learning, optimal control theory, and robotics, and include the construction of hierarchical motor skills, active learning, Bayesian optimization applied to control, and machine learning algorithms with high-probability safety and fairness guarantees.