The basic goal of our research is to investigate how humans learn and reason, and how intelligent machines might emulate them. In tasks that arise both in childhood (e.g., perceptual learning and language acquisition) and in adulthood (e.g., action understanding and analogical inference), humans often paradoxically succeed in making inferences from inadequate data. The data available are often sparse (very few examples), ambiguous (multiple possible interpretations), and noisy (low signal-to-noise ratio). How can an intelligent system cope?
We approach this basic question as it arises in both perception and higher cognition. Our research is highly interdisciplinary, integrating theories and methods from psychology, statistics, computer vision, machine learning, and computational neuroscience. The unified picture emerging from our work is that the power of human inference depends on two basic principles. First, people exploit generic priors — tacit general assumptions about the way the world works, which guide learning and inference from observed data. Second, people have a capacity to generate and manipulate structured representations— representations organized around distinct roles, such as multiple joints in motion with respect to one another in action perception. Our current areas of active study include action understanding, motion perception, object recognition, causal learning, and analogical reasoning. Below are a few examples of recent and ongoing research projects.
The ultimate goal of perception and cognition is to enable effective interactions with the external environment. Three key questions that connect action perception to understanding and reasoning are: (1) what information do humans use efficiently for different action related tasks, such as identifying individual actors, or searching for a fighting person in a crowd? (2) How do humans efficiently categorize different types of actions? (3) How do individuals interact with other people? We integrate modern modeling approaches with behavioral experiments to investigate how humans form categories for different types of actions, and how knowledge of one action type can facilitate inferences about other action types through analogical mapping. We aim to bridge perception and causal reasoning to investigate how humans understand social activities by perceiving and reasoning about actions involving interacting agents.
We study motion perception for cases ranging from simple translational motion, to complex radial/circular motion, to the sophisticated biological motion on which human actions are based. The integration of psychophysical experiments with computational models assists us in understanding (1) how humans represent different motion patterns as structural complexity increases; (2) the limitations of human motion processing; (3) the strategies the human visual system employs to complete different tasks (e.g., motion segmentation, motion grouping, and tracking over time).
What are the salient and invariant features used by the human visual system for object recognition? We address this question using synthesized images. Generative models are used to produce synthesized images that are visually realistic but with well-controlled stimulus properties. We use these stimuli to study how human recognition performance is affected by adding or deleting feature sets in the image.
We formulate causal learning within a Bayesian framework to differentiate two fundamental questions: (1) What likelihood functions do humans use in causal inference? (2) What prior knowledge do humans assume? A successful inference model incorporating a theory of both likelihoods and priors should provide coherent answers to a variety of causal queries and experimental designs. Our research aims to develop Bayesian models for a range of experiments on causal learning, and to assess the validity of computational models by comparing their predictions with human performance.
We study analogical reasoning from a computational perspective. Specific research questions include (1) What is a normative model for analogical reasoning, and how does it compare to actual human performance? (2) How are relational representations constructed from non-relational inputs, and how are they matched in analogical reasoning?