Optimization Approaches for Large-scale Pattern Recognition Tasks

Dimitri Kanevsky


Pattern recognition algorithms fall short in cases that involve large data sources, such as large vocabulary speech recognition or translation, enterprise applications like data mining and business intelligence, or analyzing fMRI data. Given the importance of all of these domains to business and to society, there is significant value in identifying novel and more effective optimization approaches to better identify patterns and improve predictive power. Over the last two decades, speech recognition technology has refined and advanced techniques to address these problems by extracting multidimensional features from a speech signal, along with statistical modeling and discriminative training techniques. IBM Research, along with university collaborators such as the Technion and Cambridge University, are now applying the knowledge gained in the speech domain to other large data challenges, through an exploratory research project. In this talk we describe how our exploratory research project attempts to solve these challenging problems in context of fMRI data analysis.