Research


I am interested in various topics in machine learning, including adversarial machine learning, learning tractable probabilistic models, and learning statistical relational models. I have also worked on desktop activity recognition, spam filtering, and recommender systems.

Adversarial Machine Learning

I spent the summer of 2004 at Microsoft Research working with Chris Meek on the problem of spam. We looked at a common technique spammers use to defeat filters: adding "good words" to their emails. We developed techniques for evaluating the robustness of spam filters, as well as a theoretical framework for the general problem of learning to defeat a classifier (Lowd and Meek, 2005ab [pdf] [pdf]). We have new results for unions and intersections of half-spaces, showing that non-linear classifiers can also be vulnerable to similar attacks (Stevens and Lowd, 2013 [pdf] [ppt]).

More recently, I have developed algorithms for learning robust models for structured prediction. CACC learns collective classification models that remain effective when some of the features are manipulated adversarially (Torkamani and Lowd, 2013 [pdf]). More generally, we showed that robustness is equivalent to regularization for structured prediction, so robust optimization can be done efficiently by constructing an appropriate regularizer (Torkamani and Lowd, 2014 [pdf] [ppt]).

Learning for Efficient Inference

Inference in Bayesian networks and Markov networks is intractable in general, but many special cases are tractable. Often, a tractable subclass such as naive Bayes mixture models yields comparable accuracy while offering exponentially faster inference (Lowd and Domingos, 2005 [pdf] [ppt] [appendix]). Furthermore, by incorporating a preference for tractable models into the learning algorithm, we can guarantee efficient inference without restricting ourselves to any particular class (Lowd and Domingos, 2008 [pdf] [pdf+proofs] [ppt]; Lowd and Rooshenas, 2013 [pdf]). Combining our methods with sum-product network (SPN) learning algorithms, we obtain the best results for SPN structure learning, often outperforming intractable Bayesian networks (Rooshenas and Lowd, 2014 [pdf] [ppt]). Given an intractable model, we can use learning methods to find an accurate but tractable approximation to the original (Lowd and Domingos, 2010 [pdf] [proofs]).

Software:

Statistical Relational Learning

Statistical relational learning seeks to represent the complexity and uncertainty present in most real-world problems by combining first-order logic with probability. The main challenges are in developing effective representations and effective algorithms. One of my projects has been Recursive Random Fields (RRFs), a multi-layer generalization of Markov logic networks that resolves a number of inconsistencies in the Markov logic representation (Lowd and Domingos, 2007a [pdf] [ppt] [ppt+audio]). I have also worked on applying quadratic optimization algorithms to Markov logic weight learning, resulting in more accurate models in much less time than before (Lowd and Domingos, 2007b [pdf] [ppt] [video]).