Qiang Liu
New. I will be an assistant professor in the Department of Computer Science at Dartmouth College starting in Summer 2015. PhD positions are available. Please email me if interested.
Workshops coorganized:
ResearchMy research area is machine learning and statistics, with interests spreading over the pipeline of data collection (mainly crowdsourcing), learning, inference, decision making, and various applications under the framework of probabilistic graphical models. Crowdsourcing. All machine learning processes start from data collection. Crowdsourcing is a modern approach to collect large amounts of labeled data by hiring anonymous workers through online platforms such as Amazon Mechanical Turk. Unfortunately, the crowdsourced workers are often unreliable and uncontrollable, raising many challenging computational questions, such as how to aggregate labels from workers with different expertise, how to combine and balance noisy (but cheap) crowdsourced labels and accurate (but expensive) expert labels, and how to crowdsource complicated objectives such as protein structures.
Learning. Learning refers to constructing probabilistic models from empirical data, either to estimate the model parameters with predefined model structures, or even to estimate the model structures solely from data? I am interested in developing efficient, possibly distributed, learning algorithms, that perform well on real world data.
Inference. With given graphical models, either handcrafted or learned from data, inference refers to answering queries, such as marginal probability (or partition function), maximum a posteriori (MAP) estimation, or marginal MAP, the hybrid of marginalization and MAP. I am interested in developing efficient inference algorithms, mostly based on variational methods and in the form of beliefpropagationlike message passing algorithms.
Structured decision making. In practice, we often need to take a sequence of actions to achieve a predefined goal, usually under uncertain environments where information is observed sequentially and interactively as we progress. Decision networks (also called influence diagrams) are graphical model style representations of such structured decision making problems under uncertainty. Just like Bayesian networks generalize Markov chains or hidden Markov chains, decision networks generalize Markov decision processes (MDP), or partially observable decision processes (POMDP). Unfortunately, the problem of finding the optimal actions for decision networks is much more challenging than answering queries on Bayesian networks, especially in cases where limited information is observed or where multiagent cooperation is required (such as in robot soccer games).
Applications. I am interested in applying these machine learning methods in many application areas.
