One of the key enabling features underlying deep learning is continuous embedding. At one end, we frame continuous embedding of objects from the point of view of metric recovery. We demonstrate that metric recovery is possible even on the basis of random walks over unweighted directed graphs, and illustrate recovery algorithms from co-occurrences in the context of word embedding. At the other end, we address the key downside arising from a pervasive use of continuous embedding within larger architectures: predictions, while generally accurate, cannot be justified in a manner suitable for human consumption or communication. I will describe work towards learning rationales in an unsupervised manner together with the supervised training of the predictor itself.
The talk covers joint work with David Alvarez, Tatsu Hashimoto, Tao Lei, Yi Sun, and Regina Barzilay.
Tommi S. Jaakkola received M.Sc. in theoretical physics from Helsinki University of Technology, Finland, and Ph.D. from MIT in computational neuroscience. He joined the MIT EECS faculty late 1998. At MIT his research has focused on many theoretical and applied aspects of machine learning and statistical inference. On the theoretical side, his work includes algorithms for large scale statistical estimation and problems that involve predominantly incomplete data sources. His applied research has concentrated around inferential questions appearing in recommender systems, computational biology, and natural language processing. He has received several awards for his publications.