Generative Similarity-based Learning

Collaborators: IDL group Sponsors: Office of Naval Research (ONR)

I study statistical learning architectures for inference based on general pairwise similarities. I am particularly interested in frameworks for learning from general pairwise similarities between complex, heterogeneous objects. For this research, I have developed similarity discriminant analysis (SDA), a generative framework for similarity-based classification. Currently, I am working on extending SDA beyond supervised classification to more general learning problems, such as transfer learning, active learning, multi-task learning, and semi-supervised learning. In a related line of work, I study flexible similarity measures that capture non-metric, possibly contradictory relational information. The goal is to capture mathematically how to measure similarities between complex, heterogeneous, and often abstract objects, and to devise principled measures of similarity that integrate well with similarity-based learning architectures. Matlab software for SDA.