Hello, Friend.

Name
Matus Telgarsky

Email
mtelgars at cs dot ucsd dot edu

Locations
2016 - (?).

2014 - 2016.   Postdoc in EECS, University of Michigan. Host: Jake Abernethy.
2013 - 2014.   Consulting Researcher at MSR NYC. Host: John Langford.
2013 - 2014.   Postdoc in Statistics, Rutgers University. Host: Tong Zhang.
2007 - 2013.   PhD in Computer Science, UCSD. Advisor: Sanjoy Dasgupta.
2004 - 2007.   BS in Computer Science & Discrete Math, CMU.
2001 - 2003.   Diploma in Violin Performance, Juilliard.

Selected papers
Representation Benefits of Deep Feedforward Networks.
Matus Telgarsky.
[arXiv]
• There exist classification problems where every shallow network needs exponentially as many nodes to match the accuracy of certain deep or recurrent networks.
Convex Risk Minimization and Conditional Probability Estimation.
Matus Telgarsky, Miroslav Dudík, Robert Schapire.
[arXiv] [short video] [poster]
• Conference on Learning Theory (COLT), 2015.
• Even when the parameter space is ill-behaved (infinite dimensional, minima don't exist, not bounded or regularized), risk minimization of certain standard losses still converges to a unique object; in the finite dimensional case, uniform convergence (generalization) holds for empirical risk minimization.
Moment-based Uniform Deviation Bounds for $$k$$-means and Friends.
Matus Telgarsky, Sanjoy Dasgupta.
[pdf] [arXiv] [poster]
• Advances in Neural Information Processing Systems (NIPS), 2013.
• Generalization bounds for $$k$$-means cost and Gaussian mixture log-likelihood for unbounded parameter sets when the data has a few bounded moments (no boundedness or further modeling assumptions needed).
Margins, Shrinkage, and Boosting.
Matus Telgarsky.
[arXiv] [video]
• International Conference on Machine Learning (ICML), 2013.
• AdaBoost, with a variety of losses, attains optimal margins by simply multiplying the step size with a small constant.
Agglomerative Bregman Clustering.
Matus Telgarsky, Sanjoy Dasgupta.
[pdf] [short video]
• International Conference on Machine Learning (ICML), 2012.
• Provides the natural algorithm, with attention to: handling degenerate clusters via smoothing, Bregman divergences for nondifferentiable convex functions, exponential families without minimality assumptions.
Full list of papers.

Miscellaneous
I used to play the violin;
I coded a screensaver, a 3-d plotting tool, and a few other things if you know where to look;
my desk is always messy;
I like scifi books, pencils, ramen, and aphex twin.