时间：2015年4月20日下午1:30—4:30

地点：计算机所楼106会议室

报告人：穆亚东博士

题目：Large-Scale Machine Learning Using Zeros and Ones

摘要：Kernel-based machine learning algorithms (such as non-linear SVM) play an important role in data analytics owing to their superior performance. However, building a kernel machine is computationally expensive. In the training stage, we need to calculate the n*n kernel matrix (n is the sample number). And for testing purpose, one needs to store all the “supporting vectors”. For large data set, both indicate high complexity.

This talk introduces my recent work on accelerating the optimization of kernel machines. The key idea is to convert the original data features into a number of zeros and ones, such that the Hamming distance between these binary bits can be used to approximate the original pairwise kernel values. This way the non-linear kernel algorithms can be efficiently solved by linear solvers, which often enjoy geometric rate of convergence. Theoretic analysis is provided for the approximation error. Several experiments on large-scale visual classification benchmarks are conducted, including one with over 1 million images.

The talk will also briefly introduce another relevant work which presents an approximate algorithm for computing large-scale min/max inner product using binary bits.

报告人简介:Dr. Yadong Mu is now a senior scientist at Multimedia Department of AT&T Labs Research. He obtained the B.Sc, B.Phil and Ph.D. degrees all from Peking University. Before joining AT&T Labs, he has ever worked at National University of Singapore, Columbia University and Huawei Hong Kong Noah's Ark Lab. Dr. Mu’s research interests are in large-scale machine learning, visual analytics topics and telecom data mining.