Thomas G. Dietterich

E-mail: tgd@cs.orst.edu Phone: +1-541-737-5559 Office: KEC 2067
PGP Public Key (Last updated June 17, 2013.)

Page Contents: Research Prospective Students Publications Talks CV Software Students and Staff Course Materials Bio Sketch Conferences

Current Research

"If you invent a breakthrough in artificial intelligence, so machines can learn," Mr. Gates responded, "that is worth 10 Microsofts." (Quoted in NY Times, Monday March 3, 2004)

The focus of my research is machine learning: How can we make computer systems that adapt and learn from their experience? How can we combine machine learning with other advances in AI to build Integrated Intelligent Systems? How can we combine human knowledge with massive data sets to expand scientific knowledge and build more useful computer applications? My laboratory combines research on machine learning and AI fundamentals with applications to problems in science and engineering.

• Scientific Projects
  • Ecosystem Informatics and Computational Sustainability: Oregon State University is a leader in combining computer science and the ecological sciences to build the new discipline of Ecosystem Informatics. Ecosystem Informatics studies methods for collecting, analyzing, and visualizing data on the structure and function of ecosystems. It is an instance of an important new direction in science: Data Exploration Science (see Jim Gray's 2003 KDD talk).

    Oregon State is also part of the NSF Expedition in Computational Sustainability jointly with Cornell University, Bowdoin College, Howard University, and the Conservation Fund. This effort seeks to develop novel computational methods to address problems in ecosystem science and sustainable management of the biosphere.

    My group is involved in many Ecosystem Informatics and Computational Sustainability activities:

    • Machine Learning for Species Distribution. One of the central goals of ecology is to understand and predict the distribution of species (including the bugs that we are studying in the Insect Identification project). Given a data set that records observations of the presence (or absence) of multiple species at multiple locations, we wish to develop models that can predict their presence/absence elsewhere. We are interested not only in static distribution models, but also in process models that capture the temporal and spatial of species distributions (e.g., bird migration, flight times of moths, return of salmon, spread of invasive species, survival of endangered species, etc.). Our species distribution team includes faculty members (Matt Betts, myself, and Weng-Keen Wong), post-docs Rebecca Hutchinson and Selina Chu, and graduate students Arwen Lettkeman and Liping Liu. We collaborate very closely with the Cornell Laboratory of Ornithology and with the DataONE Datanet. In particular, we are studying methods for dealing with the many shortcomings of the citizen science data collected by the Lab of Ornithology in their eBird project including (a) partial detection, (b) wide range of birder expertise, and (c) highly biased spatial distribution of observations.

    • BirdCast. Another special case of species distribution modeling is understanding bird migration. With the Lab of Ornithology, we are developing methods for reconstructing and predicting bird migration across North America. Our goal is to provide daily forecasts of bird migration by combining eBird reports, weather radar, acoustic monitoring of flight calls, and weather forecasts. The project web site is available here.

    • Finding Swallow Roosts from Doppler Radar. One particular instance of species distribution modeling is understanding the dynamics of tree swallow behavior. Former postdoc Dan Sheldon is developing algorithms for analyzing Doppler Radar to identify swallow roosts and understand swallow migration.

    • Approximate Optimization for Bio-Economic Models. Many sustainability applications require solving large spatio-temporal optimization problems under uncertainty. We are collaborating with economists Jo Albers and Claire Montgomery on methods for approximate solution of spatio-temporal optimization problems involving land management for wildfire control and counter-measures for controlling invasive species.

    • Project TAHMO: Deployment, Cleaning, and Analysis of Sensor Network Data. We are part of the Project TAHMO that seeks to construct and deploy a network of 20,000 hydro-meteorological stations in Africa. We are developing algorithms for sensor placement, data cleaning, recovery from damaged sensors, and analysis of the resulting data. We are building on our previous work with Ethan Dereszynski on dynamic Bayesian network models for sensor data cleaning.

    • Arthropod Identification. Our current understanding of complex ecosystems is limited by a lack of data. One particularly useful kind of data is population counts of "bugs" (small arthropods that live in soils, lakes, streams, and the ocean). The BugID project seeks to develop devices for capturing, imaging, and sorting bugs combined with general image processing/machine learning/pattern recognition tools for counting and classifying them. We hope to transform the ability of scientists to measure the health of forests, streams, and estuaries. More generally, we are interested in developing a wide range of novel instruments for expanding the quality, quantity, and spatio-temporal resolution of ecologically-relevant data. Our research also contributes to computer vision and object recognition more generally.

    • NIPS 2012 Posner Lecture: Challenges for Machine Learning in Computational Sustainability.

    • ICML 2011 Tutorial on Machine Learning in Ecology and Ecosystem Management

  • Intelligent Desktop Assistants. We have been involved in two large efforts to develop intelligent assistants for the computer desktop.

    • TaskTracer. When you come into work in the morning, you don't want to say to your computer "I want to run Word", but rather, "I want to work on my CS534 homework". In other words, you would like a user interface that was organized around your projects and activities rather than around application programs, files, folders, etc. You would also like all of your information in one place rather than scattered across the local file system, network file systems, web sites, email folders, calendar, contacts, etc. TaskTracer extends the Windows UI to provide exactly this functionality. This research is supported by Google and Intel (with previous support under the DARPA CALO project). OSU News Service story. Project Web Site.

    • CALO. The goal of the CALO project was to develop an AI personal assistant that can help you find relevant documents, prepare for meetings, keep track of what is going on during meetings, and autonomously execute tasks such as arranging travel, scheduling meetings, executing administrative workflows (e.g., purchasing and staffing), and so on. Our work on CALO focused on developing methods for integrating multiple, separately-engineered components into a single learning and reasoning system. We also prototyped a novel system that employs programming-by-demonstration to define new learning tasks for CALO to solve autonomously. We are currently editing a book describing the results of the CALO project.

  • Next Generation Phenomics. An important goal in biology is to reconstruct the tree of life. As part of the Project AVATOL team, we are developing computer vision and machine learning methods to automatically discover and score phenotype characters (features) from images of biological specimens. These scores can then be combined with other information (e.g., genetic sequences, functional measurements) to reconstruct phylogenetic trees. Phenomic information is particularly valuable for sets of closely-related species (where DNA differences may not reflect functional differences) and for extinct species known only through fossil specimens.

    The computer science challenges involve learning to score known characters, which typically include shape, texture, color, and topological features of specimens, from weakly-labeled data and discovering new characters that are shared across some taxonomic groups but not others.

  • Learning via Interaction. Statistical machine learning has focused primarily on learning from observational data. Human learning often involves learning from demonstrations, explanations, and feedback. I am interested in combining all of these methods to develop intelligent systems that can learn both autonomously and through interaction with human users and coaches. We have three efforts in this direction:
    • AI for Computer Games. The AI components of most computer games are hand-authored rule-based systems. We are studying methods for developing game AI agents via machine learning and planning. One approach is to apply reinforcement learning to automatically learn an opponent for games. Another approach is to teach the game AI by demonstrating how to play the game. A third method is to provide "coaching feedback" to the learning system. A fourth approach is to transfer knowledge learned on one game to rapidly construct an AI agent for a new game. We use the RTS games Wargus and StarCraft as our experimental platforms. Wargus is based on the Strategus game engine. StarCarft is a commercial game from Blizzard Entertainment. This work is funded by ARO under a MURI grant.

    • End-User Debugging of Learned Programs. We are starting to see end-user applications that incorporate machine learning components (e.g., adaptive spam filters, adaptive email management, adaptive user interfaces). How can we empower end users to get these learning systems to behave properly? For example, how can end users define new features, provide advice on feature relevance, and yet also understand that no learning system can be perfect? How can a learning system explain itself to the end user? Funding provided by the National Science Foundation.

• Fundamental Machine Learning and Artificial Intelligence Research
  • Machine Reading and Deep Reading. In collaboration with researchers at BBN, CMU, University of Washington, ISI, and UMass, we are studying methods for extracting knowledge from text to support inference. Our focus is on learning rules (e.g., Horn clauses) and scripts (e.g., logical hidden Markov models) from noisy and incomplete training data extracted from reading text. Funding provided by the DARPA Machine Reading and DEFT programs.

  • Anomaly Detection. An important capability for AI systems is to be able to detect when an input situation is unusual. For example, anomaly detection can allow machine learning systems to detect when an input case is very different from the training data and hence could lead to extrapolation and poor performance. Anomaly detection methods are also important for detecting novel failures in sensor networks and novel attacks on computer systems. We are developing a range of algorithms for anomaly detection under the DARPA ADAMS program.

  • Flexible Latent Variable Modeling. Many problems in machine learning require learning models of hidden ("latent") processes. Such latent variable models can be easily represented using graphical models. However, such models are typically expressed using parametric probability distributions, which limits their ability to adapt to the complexity of the process and the amount of data. Our research seeks to integrate flexible machine learning methods (such as boosted regression trees) into latent process models. Postdoc Rebecca Hutchinson and graduate student Liping Liu are developing an R package that integrates boosted regression trees into certain latent variable models common in species distribution modeling.

  • Learning Individual Models from Aggregate Data. Most data in ecology (and other fields) records information in aggregated form (e.g., population counts, census figures). Often, we wish to fit models of individual behavior using such aggregated data. One example is the problem of predicting bird migration from eBird counts. Former Postdoc Dan Sheldon has developed a new formalism, the Collective Grahical Model, that directly transforms individual models to aggregate models that can then be easily linked to aggregated data.

  • Evidence Trees. We have developed a new approach to supervised learning in which ensembles of tree classifiers are applied not to make classification decisions but to select which training data points provide evidence relevant to making a decision or prediction. This evidence can then be input to a second-level decision making process, which could be another classifier or some form of kernel density estimation. We are exploring this in the context of computer vision and the Arthropod Identification project.

  • Sample-Efficient Algorithms for Solving Spatial Markov Decision Processes. We are developing algorithms for solving MDPs in which the state consists of a landscape of patches, and each patch has its own state. This means that the state space is enormous. At each time step, an action must be specified for each patch, so the action space is also enormous. In these problems, phenomena (such as fires, infections, species spread) propagate spatially, so we cannot treat the patches as independent. Our algorithms seek exact (or bounded approximate) solutions for small problem instances, and satisfactory solutions for real-sized problem instances. Our algorithms optimize both expected (discounted) cumulative reward and also risk-sensitive reward measures. The transition probabilities for these models are specified in the form of an expensive simulator which can be invoked with a given state and action to obtain a sample of the resulting state and the reward. An important goal is to minimize the number of calls to the simulator.

  • Software Engineering Methods and Tools for Machine Learning Systems. Creating and maintaining a successful deployed machine learning system is still largely an art that requires a Ph.D. The goal of this research is to develop software engineering methods and tools for creating and maintaining deployed machine learning systems.

• Reviews, tutorials, and books. I have written several review articles and tutorials on machine learning.

My group meets on Tuesdays 3-5pm in KEC 3114. Each week, one member gives a one-hour presentation on their work, and then everyone provides 5-minute reports on their progress. There is also an AI Seminar scheduled from time to time (contact Alan Fern for details).

Information for Prospective Students

I am NOT accepting any new students or postdocs for Fall 2014. The following is advice for students applying in future years.

If you are seeking a research career in machine learning, data mining, artificial intelligence and related areas, and you have a strong background in mathematics and programming, please read my Information for Prospective Students page. To see what courses I expect my Ph.D. students to take, please see Recommended Courses for Ph.D. Students in Machine Learning.

Publications, Curriculum Vita, Software, and Data

• Publications
• Talks
• Curriculum Vita
• Downloadable Collection of Error Correcting Codes for use with the error-correcting output coding technique.
• Release 1.0 of MAXQ Hierarchical Reinforcement Learning code.
• RSW (Recurrent Sliding Window) Package for WEKA for sequential supervised learning.
• Implementation of the PCBR region detector developed as part of the BugID project.
• Implementation of TreeCRF system and supporting materials from our JMLR paper. Java implementation by Brad Block.
• STONEFLY9 Image Database.
• EPT29 Image Database.
• Starcraft Scouting dataset (for UAI 2012 paper)

Professional Service, Journals, and Book Series

• I am President Elect of the Association for the Advancement of Artificial Intelligence.
• I am a member of the Editorial Board for the Journal of Machine Learning Research, which is an electronic (and hardcopy) journal covering all areas of machine learning.
• I serve on the Board of Advisors of Neural Computing Surveys.
• I edit the MIT Press book series on Adaptive Computation and Machine Learning.
• I am past President of the International Machine Learning Society.

Entrepreneurial Activities

• I am a co-founder of Strands (formerly MyStrands; formerly MusicStrands), a recommendation company.
• I am a co-founder of Smart Desktop. Smart Desktop is now part of Decho, Inc., which is a "cloud computing" effort of EMC. Decho is commercializing technology developed as part of the TaskTracer system.
• I am a co-founder of BigML. The goal of this startup is to develop large scale cloud-based machine learning services.

Current Students and Staff

• Andrew Emmott, Graduate Student.
• Selina Chu, Postdoc.
• Mark Crowley, Postdoc.
• Rebecca Hutchinson, Postdoc.
• Jesse Hostetler, Graduate Student.
• Jed Irvine, Software Developer.
• Arwen Lettkeman, Graduate Student.
• Liping Liu, Graduate Student.
• Sean McGregor, Graduate Student.
• Michael Slater, Project Manager.
• Majid Alkaee Taleghan, Graduate Student.
• Shahed Sorower, Graduate Student.
• Pat Sullivan, Assistant and Grants Coordinator.

Former Students and Staff

• Hussein Almuallim, Oil and Energy Professional, Calgary, Canada.
• Eric Altendorf, Google.
• Adam Ashenfelter, BigML, Inc., Corvallis, Oregon.
• Ghulum Bakiri, Department of Computer Science, Bahrain University
• Christian Baumberger. Master Student in Biomedical Engineering at University of Bern.
• Xinlong Bao. Google Pittsburgh.
• Brian Breck.
• Waranun Bunjongsat.
• Giuseppe Cerbone. Independent Information Services Professional, Milan, Italy.
• Martha Chamberlin.
• Hei Chan.
• Richard Charon.
• Eric Chown, Full Professor, Bowdoin College.
• Dan Corpron
• Diane Damon, Damon Consulting, Portland, OR.
• Ethan Dereszynski, Research Scientist, WebTrends, Portland, OR.
• Phuoc Do, Wiavia.
• Nicholas Flann Associate Professor, Utah State University
• Greg Foltz.
• Dan Forrest.
• Tony Fountain, Scientist
• Ashit Gandhi, Founder and Vice-President, Prism Gem, LLC - The Art of Diamond Coloring.
• Colin Gerety, Fort Collins, CO.
• Brandon Harvey, Enterprise Systems Integrator at University of Oregon.
• Guohua Hao, Data Scientist at Dataminr.
• Hermann Hild, President, SMI Cognitive Software GmbH .
• Saket Joshi, Postdoc (CI Fellow), Oregon State University.
• Varad Joshi, Senior Software Engineering Manager at Arris, Portland, OR.
• Caroline Koff, Hewlett-Packard Corporation, Fort Collins, CO.
• Victoria Keiser, Research Programmer, CMU. Masters Thesis (PDF).
• Michael Kelm, Research Scientist, Siemens Healthcare.
• Eun Bae Kong, Professor, Computer Science, Chungnam National University, South Korea
• Bill Langford, Research Associate at RMIT, Melbourne, Australia.
• Junyuan Lin, VMWare, Seattle.
• Dragos Margineantu, The Boeing Company.
• Gonzalo Martinez, Assistant Professor, Autonomous University of Madrid.
• Prafulla Mishra, Software Development Manager at eBay.
• Avis Ng.
• Soumya Ray, Assistant Professor, Case-Western Reserve University.
• Angelo Restificar, Principal Machine Learning Engineer, EBay, San Jose.
• Ritchey Ruff, Senior SDET, Microsoft.
• Dan Sheldon, Assistant Professor, University of Massachusetts, Amherst.
• Jianqiang Shen. Research Scientist, PARC. Doctoral dissertation.
• Rongkun Shen. Post-doc, Oregon Health and Science University, Portland.
• Michael Shindler, HULU.
• Shriprakash Sinha. Ph.D. student TU Delft.
• Simone Stumpf. Senior Lecturer, City University London.
• Tao Sun, Graduate Student at UMass Amherst.
• Dan Vega, Senior Software Engineer at Valley Inception, LLC.
• Mark Vulfson. Microsoft Corporation.
• Kiri Wagstaff, Researcher at JPL.
• Xin Wang, Senior Scientist at Intelius.
• Dietrich Wettschereck. Recommind.com.
• Pengcheng Wu.
• Michael Wynkoop, Qualcomm.
• Qing Yao, College of Informatics and Electronics. Zhejiang Sci-Tech University. Hangzhou, China.
• Wei Zhang, The Boeing Company.
• Wei Zhang. Google. Doctoral Dissertation (PDF).
• Valentina Zubek, Boehringer Ingelheim.

Previous Courses and Courseware

• CS519/GEO599: Principles of Ecosystem Informatics, 2004-2005.
• CS 534, Spring 2005, Machine Learning.
• CS430, Fall 2003, Introduction to Artificial Intelligence
• CS539, Fall 2003, Seminar: Probabilistic Relational Models
• CS 533, Applied Artificial Intelligence for Engineeers.
• CS 539, Winter 2000, Selected Topics in Artificial Intelligence: Probabilistic Agents
• CS 430/530, Fall 1999, Artificial Intelligence Programming Techniques.
• CS 519, Fall 1996. Research Methods in Computer Science.
• CS 450/550, Winter 1996, Introduction to Computer Graphics.

Machine Learning Resources

• How to be a Graduate Student. A great web page with pointers to lots of good resources for graduate students.
• Standard Proofreading Symbols that I use when marking corrections on papers.
• Computing Research Repository (part of arXiv). I am the curator for machine learning.
• Research Index at Penn State. An invaluable resource for finding online articles, citations, etc.
• Journal of Machine Learning Research (JMLR). (Free electronic and hardcopy journal.)
• Machine Learning. (Expensive hardcopy journal.)
• Journal of Artificial Intelligence Research (JAIR). (Free electronic and hardcopy journal.)
• The Machine Learning Database Repository at UC Irvine.
• The Machine Learning Programs Repository at UC Irvine.
• UC Irvine ML information.
• StatLib containing data, algorithms, and other information relevant to statistics.
• Knowledge Discovery Mine containing information about knowledge discovery in databases.
• NIPS Online Proceedings.
• CMU Reinforcement learning group.
• Bibliographies on Artificial Intelligence
• Computing Research Repository (CoRR) CS Preprint Service and Archive (part of arXiv).
• The DBLP Computer Science Bibliography.
• Numerical Recipes Homepage.

My Family's Musical Activities

• yOya on MySpace and yOya home page. My son Noah writes songs and plays keyboards for this band.

• Jubilate: The Women's Choir of Corvallis. My wife Carol sings in this choir.

Tom Dietterich, tgd@cs.orst.edu