Login
Computer Science @ Rochester
Monday, March 04, 2013
10:45 AM
CSB 209
Parikshit Ram
Georgia Institute of Technology
Fast similarity search for the big data era: Boosting performance with machine learning
Similarity search is ubiquitous; popular web applications such as document/image/music/movie search as well as scientific applications such as protein sequence matching are some examples. The sheer size of data necessitates efficient search algorithms. I will present my research on novel fast and scalable search algorithms that go beyond structured vectorized data or distance metrics and leverage machine learning to improve search performance.

First, I will demonstrate how provably efficient exact similarity search can be performed in the absence of vectorized data. I utilize kernel functions from machine learning to abstractly index the data and present a novel branch-and-bound algorithm. This method has a theoretical search-time guarantee logarithmic in the size of the data and significantly boosts search speed to up to 100 times faster than usual exhaustive search over real world datasets.

Second, I will describe novel formulations for approximate similarity search that are meaningful and intuitive in various applications. For example, one proposed formulation enables algorithms to provide the best approximate results given any time-budget. This allows time-sensitive applications to provide users with "good-enough" results to work on, rather than keeping them waiting. I will discuss simple yet scalable algorithms to achieve these new search-approximations.

Finally, I will present my research that takes advantage of the abundance of data with machine learning techniques to improve search performance, such as using the concept of "large margin separation" from machine learning to theoretically and empirically improve search performance.

Bio: Parikshit Ram is a PhD student at Georgia Tech in the School of Computational Science and Engineering, College of Computing. He is a member of the FASTlab which develops fundamental algorithms and statistical tools for machine learning and data mining. Pari joined Georgia Tech in 2007 after completing his BS and MS in Mathematics and Computing in the Department of Mathematics at the Indian Institute of Technology, Kharagpur.

Pari's work focuses on developing fast algorithms for similarity search and related problems. His research covers both empirical and theoretical aspects of search algorithms and develops the area of learning to search - using machine learning techniques to enhance search. Pari has published his work at various top conferences in machine learning and data mining such as NIPS, KDD and SIAM-DM and is actively involved in the development of the MLPACK machine learning C++ library.