18th Simon Stevin Lecture on Optimization in Engineering
"Convex Optimization for Machine Learning with Multi-Modal Data"
Gert Lanckriet
Department of Electrical and Computer Engineering
University of California, San Diego
http://www.ece.ucsd.edu/~glanckriet/
flyer,
poster,
slidesAbstract:
Whether it is for mining MySpace for "romantic jazz music with deep male vocals'', predicting protein function from biological data, or modeling financial markets, an increasing number of data mining and machine learning applications faces data that exhibits a variety of modalities. For example, music databases derived from crawling online information may contain audio content, as well as a variety of other modalities, like song lyrics, artist images, video clips, music reviews, billboard rankings, etc. Financial markets may be modeled based on stock quotes, as well as twitter feeds, news articles, etc. Jointly modeling these heterogeneous data sources is expected to improve the predictive accuracy of the resulting model, compared to modeling a single source of information only. For example, music recommendation could benefit from computing song or artist similarity based on all available modalities, instead of using audio content alone.
By encoding each modality as a kernel (i.e., similarity) function, learning from multi-modal data can be cast as a multiple kernel learning (MKL) problem. MKL optimally integrates heterogeneous data into a single, unified similarity space. Optimality depends on the specific machine learning problem at hand (binary or multi-class classification, ranking, visualization, etc.). This talk will leverage convex optimization to formulate MKL approaches for various machine learning algorithms, going from the original MKL formulation that learns a weighted linear combination of kernels, to a novel technique that learns a concatenation of linear projections, where each projection extracts the relevant information from a kernel's feature space. The new formulation results in a more flexible and more general model than the original formulation. The various MKL formulations will be illustrated for different algorithms and applications.
Biographical Information
Gert Lanckriet received a Master's degree in Electrical Engineering from the Katholieke Universiteit Leuven, Leuven, Belgium, in 2000 and the M.S. and Ph.D. degrees in Electrical Engineering and Computer Science from the University of California, Berkeley in 2001 respectively 2005. In 2005, he joined the Department of Electrical and Computer Engineering at the University of California, San Diego, where he heads the Computer Audition Lab (CALab). He was awarded the SIAM Optimization Prize in 2008 and is the recipient of a Hellman Fellowship, an IBM Faculty Award, an NSF CAREER Award and an Alfred P. Sloan Foundation Research Fellowship. His lab received a Yahoo! Key Scientific Challenges Award, a Qualcomm Innovation Fellowship and a Google Research Award. His research focuses on the interplay of convex optimization, machine learning and applied statistics, with applications in computer audition and music information retrieval.
About the Lecture Series:
The "Simon Stevin Lecture Series on Optimization in Engineering" is set up in order to promote optimization in engineering. For this aim, every quarter of the year an outstanding international scholar is invited to report on latest progress in the development of optimization algorithms and their applications in engineering.
Simon Stevin (1548-1620) was a Flemish mathematician and engineer. Among other, he helped to advance the use of decimal fractions, was the first to explain the tides by the attraction of the moon, and discovered the hydrostatic paradox. He made numerous inventions, among them a wind propelled carriage with sails, the "land yacht", which once impressed Prince Maurice of Orange as it moved faster than horses, in around 1600 on the beach between Scheveningen and Petten. Simon Stevin was fond of promoting the use of science in daily life and in craftmanship, and translated various mathematical terms into dutch. Among other, he introduced the dutch word for mathematics, "wiskunde".
