CAREER: Discriminative and Generative Machine Learning with Applications in Tracking and Gesture Recogniton

Columbia University, New York NY

Investigators

Abstract

This project aims to develop a tighter integration of generative (e.g., Bayesian networks) and discriminative (e.g., support vector machines) machine learning. These two types of learning are typically not integrated in current practice and are often seen as competing approaches; however, such integration is crucial in complex multi-disciplinary domains, such as vision, speech, and computational biology, where scientists have real expertise and exploitable knowledge about elaborate systems yet also need machine learning to achieve optimal performance for specific tasks. This project's integrated generative-discriminative framework will allow practitioners to flexibly design and structure a given learning problem using generative tools like Bayesian networks and then to maximize the performance of these models using discriminative methods like maximum entropy and probabilistic kernels. This framework will be used in computer vision tracking applications and in classification of gestures. In a laparoscopic robotic surgery platform, the methods will be used for classifying surgical drill movements and predicting surgeon dexterity level. The project's unified approach will be used to create a more comprehensive machine learning course experience for students, complete with online class materials, visual demonstrations and software toolkits.