Daniel Hsu develops algorithms for statistical analysis and machine learning. He focuses especially on settings that involve high-dimensional data or interaction. His research has produced the first computationally efficient algorithms for several statistical estimation tasks, provided new algorithmic frameworks for solving interactive machine learning problems, and has led to the creation of scalable tools for machine learning applications.
Statistical models posit the presence of interesting structure in data; the goal of estimation is to measure and quantify this structure. Estimating parameters of statistical models is especially challenging when some of the variables in the model are not observed (such as in mixture models and hidden Markov models). Indeed, classical methods often require solving intractable optimization problems (e.g., maximum likelihood estimation), and therefore are challenging to use in applications involving large, high-dimensional data sets. Hsu develops and analyzes computationally efficient algorithms for estimation in hidden variable models (including mixture models, hidden Markov models, and topic models), as well as for other estimation problems where spectral analysis is especially important. Interactive machine learning concerns learning agents that adaptively make decisions that ultimately affect the data available to the agent. For example, the agent may be the learning procedure used by a data scientist, and it may interact with the data scientist to jointly construct an accurate classifier. Such scenarios extend beyond the standard frameworks for understanding machine learning algorithms. Hsu's research has developed new frameworks, algorithms, and analytic techniques for interactive learning.