Social and Embodied Intelligence
Beyond Chatbots
Human experience is nuanced, social, and embodied. This kind of experience, and associated abilities, are missing in current AI systems. To interact successfully with humans in human-centered contexts, AI systems need to learn and use “world models” that enable context-sensitive social and embodied reasoning.
Goals
- Develop AI agents with situated, multimodal world models
- Propose new architectures, learning objectives, data distributions, and combinations thereof that induce agents to learn causal models of their environments
- Advance scientific understanding of how human capacity for embodied and social cognition originates and develops
Current Projects
Building multimodal world models
Determine the best methods for endowing AI agents with integrative multimodal world models that integrate text, image/video, and time-series information into a single coherent representation personalized to a specific user and context.
Perspective taking in the 3D physical world
Develop datasets, learning algorithms, and embodied evaluations that produce AI systems with human-like abilities for visual perspective-taking.
Learning from intrinsic motivation
Develop active-learning and reasoning methods that incorporate human-like intrinsic motivations (curiosity, drive for causal understanding) and inductive biases (parsimony, abstraction), evaluated in idealized and real-world environments.