Wednesday, October 2 at 4:15pm
Packard Laboratory, 466
19 Memorial Dr W, Bethlehem, PA 18015
Dr. Lizhong Zheng, from MIT, will present on "Universal Features--Information Extraction and Data-Knowledge Integration"
Abstract: With the growing demand of using data analytics in a wide range of applications, a key research challenge has emerged to represent data in a generic semantic space, where we need to have a quantitative way to represent the useful information and knowledge succinctly, and at an abstract level. The key issues include how to define a universal interface for knowledge representation, how to manage and integrate the knowledge from multiple data sources, how to utilize domain knowledge, and how to cope with non-ideal situations such as the disparity in the quality of different datasets and precision losses in the processing. There are numerous algorithms as possible ways to achieve such goals. Particularly, neural networks are expected to play a key role. The main difficulty is that we still do not have a complete theory about deep learning, to identify exactly what knowledge is learned by neural networks, what hidden assumptions are needed for the desirable performance. In this talk, we try to address this problem by developing a theoretical structure to measure the meaning of information by its relevance to specific inference problems, and from that we explain the behavior of neural networks as extracting “universal features”, defined as the solution to a specific optimization problem. This helps us not only to understand the learning process inside a large neural network, but also to draw connections to a number of well-known concepts in statistics and other learning algorithms. Based on this theoretic framework, our goal is to develop more flexible, robust, and interpretable data embedding algorithms.