BIDMach - an interactive, general machine learning toolkit for Big Data

Its a data-centered world now, and machine learning is the key to getting value from data. But we believe much of the value from Big Data is untapped, and requires better tools that are much faster, more agile and more tunable (allowing tailoring of models). The current wave of tools rely primarily on cluster computing for scale-up. The BID Data project focuses on single-node performance first and fully taps the latest hardware developments in graphics processors. It turns out this approach is faster in absolute terms for most problems (i.e. our tool on a graphics processor outperforms all cluster implementations on up to several hundred nodes), is fully interactive and supports direct prototype-to-production migration (no recoding). Some problems (e.g. training large deep learning networks), still benefit from scale-up on a cluster. We have developed a new family of communication primitives for large-scale ML which are provably close-to-optimal for a broad range of problems, and e.g. they hold the current record for distributed pagerank. Our most recent work is on live tuning and tailoring of models during optimization, and we have developed a new approach to optmization: parameter-cooled Gibbs sampling to support this.

Speaker Bio:

John Canny (http://en.wikipedia.org/wiki/John_Canny) is a professor in computer science at UC Berkeley. He is an ACM dissertation award winner and a Packard Fellow. He is currently a Data Science Senior Fellow in Berkeley's new Institute for Data Science and holds a INRIA (France) International Chair. Since 2002, he has been developing and deploying large-scale behavioral modeling systems. He designed and protyped production systems for Overstock.com, Yahoo, Ebay, and Quantcast. He currently works on several applications of data mining for human learning (MOOCs and early language learning), health and well-being, and applications in the sciences.