People excel at real-life tasks that remain challenging for modern Artificial Intelligence (AI), such as learning new tools from a few examples, creative thinking, and reusing skills to solve new problems. This data-efficiency and broad generalization remain some of the most salient gaps between human and machine intelligence. Although people rarely adhere to normative economic theories, their behavior is often resource-rational -- optimally balancing computational costs against the value of computation. In the context of real-world reasoning, resource-rationality entails developing functional and resource-efficient models of the world and using these models to plan and reason in flexible and generalizable ways. We have three fully funded PhD positions that will explore different approaches to building efficient AI models inspired by human cognition, with the goal of promoting our understanding of cognitive psychology and advancing efficient human-like AI.
Our lab conducts interdisciplinary research combining methods from AI, neuroscience, and psychology in new ways to develop improved state-of-the-art models of human cognition and advance AI. You will work with an open-minded group of researchers at various stages of their careers in a cooperative, collegial, and interdisciplinary lab. We value and actively engage in national, North American, and international collaborations. You will join an interdisciplinary collaborative network that includes leading experts in artificial intelligence, program synthesis, Large Language Models, architecture, and psychology.
Current AI models formalize planning as a search in a decision tree of potential actions and outcomes. The size of this tree determines the computational cost of the problem, or its theoretical difficulty. However, theoretical difficulty rarely aligns with human experience, as people often easily solve problems deemed intractable in theory. The key characteristic of real-world problems that people are so adept at is their compositional, semi-regular structure, such as predictable patterns of hills and valleys in nature, or structured street networks in built environments. Computational study of how people plan in such contexts is central to engineering resource-efficient AI that can plan with human-like efficiency, while adapting to increasing problem scales. In this project, we will integrate AI, Large Language Models, and psychology in new ways to develop a computational understanding of how people conceptualize maps and plan in realistic spatial navigation tasks. We will build computational models that can anticipate environmental structure and create human-like AI that can produce efficient plans in realistic domains.
Existing cognitive models of planning (e.g., in games like chess) tend to pre-specify possible planning models as anchored to classic algorithms, such as MDP solvers and stochastic search. In contrast, people likely maintain and learn an evolving library of planning strategies encoded as mental programs, and grow this library through experience. In this project, we explore approaches to modeling how such libraries of programs evolve and grow through social interaction and experience. We will work in collaboration with cognitive and developmental psychologists to build AI that can grow and learn like a child.
In recent years, we have seen impressive growth in Creative AI, such as diffusion models that can generate images and GPT models that can tell stories. However, such models are still difficult to customize to individual preferences and are prone to repeating generic artistic styles, while most application use-cases require that an AI assistant can become an AI apprentice, learning from and co-creating with the artist. In this project, we will explore neuro-symbolic human-like AI models of the human creative process that can understand and mimic the human cognitive approach to creation by learning libraries of cognitive concepts and recombining these concepts in new ways. We will specifically apply these models to musical composition, building an interactive AI accompanist that learns to play alongside a musician to interactively accompany a performer in an improvised musical jam session. This project will be co-supervised by Sageev Oore and will use a physical piano (disklavier). Understanding musical production, notation, and harmony (at a minimum, a self-taught knowledge of classical harmony) is important to this project. Students with a background in musical theory and experience with playing a musical instrument are encouraged to apply.
Contact us if you are a prospective graduate student interested in the above, or similar, research projects. If you have your own project idea related to modelling how people plan, explore or search for information, feel free to reach out and I will be happy to discuss it with you.
Please include in your email:
[under construction]
[under construction]