Research Interests

This page gives a brief overview of my research interests. Links to relevant publications can be found here.

My primary area of interest in research concerns computational theories of inference and learning in biological and artificial systems. Questions that motivate me include: How can we construct artificial systems that can analyze complex, cluttered environments with the apparent ease and accuracy of natural systems? How can novel visual items be processed efficiently, and how do the representations and processing mechanisms change as items become more familiar? Below I briefly describe three inter-related areas that I am currently focusing on in my research, in collaboration with a number of students and other investigators.

Machine Learning

We develop methods for computers to learn from data. These models are intended to not only extract useful patterns from large sets of data, but also to perform well in practice by inferring how these patterns apply to unseen data. The models we pursue are formalized based on probability and information theory, and learning and inference are related aspects of the approach -- learning involves adapting the models, while inference or perception entails inverting the models.

Our current research focuses on a couple issues: How can these models be used for unsupervised learning, in which the aim is to develop useful internal representations of data in situations in which correct labels are not provided? Are there advantages to be gained from a greedy sequential approach to training such systems, and how do recent successful ensemble methods, such as boosting, relate to these probabilistic learning models?

Neural Computation

On the biological side, we aim to formulate and analyze computation in networks of artificial neurons that can potentially lend insight into neural and cognitive processing. Our research concerns representations in the brain, and primarily focuses on population codes, in which a set of units collectively encode information about a relevant variable in the environment, such as spatial position or motion direction. This form of representation is fault-tolerant and efficient, and appears to be ubiquitous in the brain. The analysis and utility of population codes is currently a central issue in neural computation.

We study many aspects of population codes, including novel learning and inference algorithms for population codes; how these codes can be used to store information beyond a single value of the relevant variable(s), such as uncertainty and more than one value; and how information in several different codes can be combined. We also apply these information processing conceptions to gain some understanding of neural computation, by using them to account for data collected from neuroscience experiments.

Perceptual Learning and Attention

We also investigate learning and vision issues in a range of perceptual experiments with human subjects. These experiments have focused on questions such as the role of experience in shaping visual representations; how occlusion affects how objects are perceived and attended to; and the degree to which memories of novel objects are specific to the location in which they are learned.

Finally, we formulate computational approaches that are useful for examining cognitive processing. We have developed a statistical formulation of attractor networks, a form of associative memory often used to model cognitive processing. We have also studied adaptive methods of segmenting images, that is, deciding which image features belong to which objects, based on statistical regularities extracted from a set of training images.