Learning Hybrid Models For Planning
Plans live a double life. On the one hand, they are discrete mental representations that can be productively recombined; on the other hand, they specify continuous motor actions that generalize across noise and environmental variation. Children quickly learn to make plans that bridge these two levels, executing multi-step actions and using tools proficiently to achieve their goals.
This talk will address the problem of learning such hybrid models. Drawing on the literature on task and motion planning in robotics, we will introduce a model that learns, from a few demonstrations of continuous actions, a set of high-level kinematic constraint representations that afford planning. We show that these representations generalize to different environments and that they can be productively recombined. We end by discussing current work on allowing these planning representations to also reason about dynamics.
Understanding language with grounded program induction
Language is powerful. Against benchmarks for machines, our own natural languages offer what we demand, or dream of demanding, from our best computational interfaces. They are flexible and adaptable: we can combine a finite vocabulary to describe an infinitude of meanings, coin new words to bend our language to changing domains, and apply abstract descriptions of the world to a multitude of vast and varying scenes. This talk will discuss language learning models that draw on the toolsets of formal linguistics and semantic parsing to frame natural language understanding as a problem of program induction, using the syntax of language to build up complex representations of meaning from compositional semantic primitives. We will present ongoing work applying semantic parsing to three disparate tasks: guiding domain-specific concept acquisition for program induction itself, learning and generalizing to highly compositional synthetic descriptions of events, and learning verb meanings from grounded natural language.
A New Frontier in Clinical Neuroscience
Veterans returning from war may continue to experience cognitive and social difficulties long after leaving the combat zone. Over the years, various generations have described these symptoms of war in various ways, "shell-shock" or "combat fatigue" for example, but regardless of the label, often times veterans must simply carry the burden of their experiences for the remainder of their lives. The current statistics on veteran suicide are an alarming indicator of the severity of the problem. Perhaps it is now time to look outside of traditional approaches to mental health treatment and explore alternative methods though the rigor of modern science. In my talk, I will discuss some possible methods to investigate the effectiveness of psychedelic drug-assisted therapies for mental health issues common to combat veterans.