A new research project on Representation Learning will start from this year, and will be sponsored by both US Office of Naval Research Global & US Air Force Office of Scientific Research.
Abstract: The goal in this project is to develop an effective framework for learning a representation of the hidden semantics from massive text collections. Unlike existing approaches, our framework should be able to deal with different forms of semantics, such as word meanings, themes, entity interaction, emotions, trends, social communities, etc. Many challenges have to be resolved, such as the non-convexity of the learning problem, the unstructured nature of text, the heterogeneity of texts, links, images, tags, etc. Nonetheless, a good way to represent and learn the hidden semantics will have a significant impact in many areas. To this end, our methodology will base on topic modeling, neural networks, manifold learning, and stochastic optimization. The framework will be employed in practical applications including text modeling, sentiment analysis, recommender systems, social network analysis.