Nicholas Haber, Postdoctoral Research Fellow, NeuroAILab, Department of Psychology, Stanford University
Title: Curiosity, active learning, and animate attention
Abstract: In everyday experience, our environments are filled with a wide variety of stimuli, some easy to understand and predict, some more challenging, and some effectively impossible to anticipate precisely. How do we engineer an artificial system to be able to explore a real-world environment so as to learn as much as possible about its surroundings, as efficiently as possible? In particular, what if the environment contains animate stimuli? Within the wide variety of stimuli we see every day, animate ones occupy a special position: they are in some ways very unpredictable, yet they also have very rich, understandable behaviors. I will describe a computational approach for artificial systems that, by virtue of curiosity (formalized and implemented through deep reinforcement learning), decide what is interesting in their surroundings, and in doing so, attend to animate stimuli and better understand them. Lastly, I will describe human subject experiments for which we hope to bring to bear these computational approaches as models of human behavior.