My research is driven by the simple desire to make computers smarter. I work on statistical machine learning and its applications. This page contains a sample of representative research projects from my group. For a complete list of projects, data sets, and code, please visit my publications page. My group has been supported by the National Science Foundation, AFOSR, and University of Wisconsin-Madison.

Semi-supervised learning

Semi-supervised learning has been widely regarded as an attractive alternative to supervised learning because of the potential savings in labeling cost. We made several contributions in the past: harmonic function and Gaussian random fields, multi-manifold learning, and matrix completion for semi-supervised learning. However, recently people realized that under certain conditions semi-supervised learning can lead to inferior performance. We are investigating safe semi-supervised learning to ensure that it is no worse than supervised learning. Another direction is online semi-supervised learning and its synergy with other learning paradigms such as active learning. [project website]

Machine learning for human learning

We are seeking the mathematical principles underlying both machine learning and human learning. This project spans a wide spectrum of research topics ranging from deriving computational learning theory to conducting human behavioral experiments. I am leading the effort with a group of professors from computer science, psychology, educational psychology, and electrical and computer engineering. We have made a number of interesting observations in human semi-supervised learning, test item effects, human manifold learning, human active learning, Rademacher complexity of the human mind, and using co-training as a human collaboration policy. This project helps us understand how humans learn and inspires new machine learning algorithms. [project website]

Text-to-picture synthesis

This project develops algorithms that automatically generate pictures from natural language sentences so that the picture conveys the main meaning of the text. Our approach is based on statistical machine learning and integrates semantic role labeling, image selection, and scene composition. With professors from psychology, we have demonstrated its effectiveness in improving young student's reading comprehension and math skills. We are collaborating with the Department of Communicative Disorders to develop an iPad app as an augmentative and alternative communication device for people with disabilities (right). A prototype is being used by clinicians in therapy for patients with autism. [project website]

Domain knowledge + latent Dirichlet allocation

One way to let human intelligence help machine learning is to incorporate the former into a probabilistic model as a prior. This is particularly useful for unsupervised learning models such as Latent Dirichlet Allocation, which otherwise learn solely from data statistics. We are developing a series of general purpose models for domain experts, with the goal to facilitate the application of topic modeling in many scientific domains. These models range from simple knowledge like "this word must be in one of these topics," to ΔLDA for statistical software debugging, to Dirichlet forest, to fold.all which allows arbitrary complex domain knowledge specified in First-Order Logic and enforced via stochastic gradient descent. Here is a summary presentation.

Software provenance recovery

Each computer program undergoes a variety of transformations: instantiation in a programming language, compilation by particular compiler with specific compilation options, and possible post-compilation transformation like link-time optimization or obfuscation or packing. Given a program binary without any symbolic information, we use machine learning techniques (CRFs) to recover its provenance -- the specific stages of the transformative process, including authorship, source language, compiler and optimization levels, and infer function entry points. Software provenance gives insight into how binaries were produced, with applications in software engineering and debugging, malware analysis, and digital forensics.

Unconventional text classification: wishes, metaphors, creativity

Conventional text classification was about grouping documents by topic. There is increasing interest in classifying text along novel dimensions. We built classifiers to identify wishful expressions, recognize the use of metaphors, gauge writer creativity, and analyze sentiment polarity in text documents. The figure on the right shows a break down of new year's wishes in 2008.