Martin Szummer's home page

I am a researcher at the University of Cambridge, UK. I am part of the Spoken Dialogue Systems group. Previously, I was a researcher at Microsoft Research in Cambridge, in the machine learning and perception group. I obtained my Ph.D. in machine learning at MIT in 2002, and my M.S. degree at the MIT Media lab.

I develop probabilistic models that describe the structure of data, for example how the pattern of words in a normal email message differs from that of a spam message, or how the pixels in an image are arranged into objects.

Research interests

•  Deep learning and flexible models. Data is often complex and our understanding of it is limited. I research deep learning models that can learn complex structure in the data. These include deep belief networks and deep auto-encoders.
•  Learning how to improve interactions with users. Via reinforcement learning, we explore interaction strategies, and gradually exploit the good ones we find.
•  Learned semantics of natural langage. We learn from user dialogues (Spoken Dialogue Systems) and from user queries and clicks in search engines.
•  Methods that control overfitting. I develop Bayesian inference techniques, which do not overfit, because they do not aim to fit - instead they integrate out model parameters. I also work on regularizers that control smoothness by exploiting partially labeled data. Such semi-supervised learning is applicable when large amounts of unlabeled data can be gathered easily, but where we do not have enough human resources to manually label all of the data.
•  Structured prediction. When making predictions over multiple items, we must model the correlations and interactions between items, in order to make predictions that are consistent across the items. For example, we may need to rank a set of items from best to worst, or classify multiple items (e.g. pixels in an image, a random field).
•  Probabilistic programming. We lift the level of abstraction to program in terms of probabilistic models, which allows significantly more complex models to be intuitively expressed and correctly implemented.
•  Big Data. I work with web-scale datasets, such as records from tens of millions of users, which I process on cloud-based clusters of 10,000 machines, using map-reduce.

Applications

• Text mining and language understanding. Learning the meaning of phrases from user interactions, data mining of user clicks.
• Image recognition, image search, handwriting recognition
• User behavior modeling from clicks, browsing data and computational advertising.

Publications

list (without abstracts) and list (with abstracts).

Selected recent publications:

• M. Gašić, C. Breslin, M. Henderson, D. Kim, M. Szummer, B. Thomson, P. Tsiakoulis, S. Young. POMDP-based dialogue manager adaptation to extended domains SIGDIAL 2013 Best Paper Award.
• M. Szummer, M. Henderson, C. Breslin, M. Gašić, D. Kim, B. Thomson, P. Tsiakoulis, S. Young. The BUDS POMDP Dialogue System Advances in Neural Information Processing Systems (NIPS) Demo 2012.
• Martin Szummer, Emine Yilmaz Semi-supervised Learning to Rank with Preference Regularization October 2011. Conf. Information and Knowledge Management (CIKM) Poster
• Daniel Sheldon, Milad Shokouhi, Martin Szummer, Nick Craswell LambdaMerge: Merging the Results of Query Reformulations February 2011. Web Search and Data Mining (WSDM) Poster
• Martin Szummer, Pushmeet Kohli, Derek Hoiem. Learning Random Fields using Graph Cuts. October 2010. Book chapter in book on MRFs, MIT press, edited by Andrew Blake, Carsten Rother, Pushmeet Kohli.
• Marc'Aurelio Ranzato, Martin Szummer. Semi-supervised Learning of Compact Document Representations with Deep Networks July 2008. Proc. Intl. Conf. on Machine Learning (ICML) 2008 792-799
• Lorenzo Torresani, Martin Szummer, Andrew Fitzgibbon. Learning Query-dependent Prefilters for Scalable Image Retrieval ( supplement ) June 2009. Proc. Comp. Vision Pattern Recogn. (CVPR)
• Nick Craswell, Martin Szummer. Random Walks on the Click Graph July 2007. SIGIR Conf. Research and Development in Information Retrieval 239-246

Datasets

Students

I have had the privelege to recruit and mentor talented students, including

• Percy Liang, now assistant professor at Stanford
• Yuan Qi, now assistant professor at Purdue
• Balaji Krishnapuram, now senior manager R&D Siemens
• Phil Cowans, now CTO Songkick.com
• Andriy Mnih, now Google DeepMind
• Marc-Aurelio Ranzato, now Researcher, Facebook
• Roger Grosse, now Ph.D. student at MIT
• Volodymyr Mnih, now Google DeepMind
• Alex Spengler, now Microsoft Research

Personal

My email address is my last name at media.mit.edu