Discriminative Classification with Incomplete Data - Lecture

Explore a comprehensive lecture on discriminative classification methods for structured data with incomplete information. Delve into the challenges of classifying sequences in various fields such as speech recognition, image and text classification, and computational biology. Learn about a new framework based on generalizations of the maximum entropy principle, designed to address issues like partially known examples, shortage of labeled training data, and abstract category labels. Discover how this approach accommodates uncertain or missing labels, extends to anomaly detection, and incorporates large margin classification with a Bayesian perspective. Gain insights into the technical aspects, experimental results, and current limitations of this method. Understand how this framework subsumes other standard discriminative methods like support vector machines. The lecture, delivered by MIT's Tommi S. Jaakkola, draws from joint work with Marina Meila and Tony Jebara, offering valuable knowledge for researchers and practitioners in machine learning and statistical inference.

Discriminative Classification with Incomplete Data – Tommi S. Jaakkola (MIT)