Artificial Intelligence

Deep Neural Networks for Structured Prediction

Thursday, March 05, 2020, 11:00am - 12:00pm PDTiCal
This event is open to the public.
AI Recruitment Talk
Amirmohammad "Pedram" Rooshenas, PhD

Recent advances in deep neural networks (DNNs) have revolutionized fields such as natural language processing, computer vision, and robotics, while also recently impacting other fields such as biology,  computational mechanics, and health care. Many of the problems appearing in these domains are structured prediction tasks, which involve predicting multiple output variables whose correlations jointly form a structure  (e.g., a set, sequence, tree, or arbitrary graph). Classically, these correlations were modeled using graphical models (such as conditional random fields) defined over the output variables. In this talk, I will describe how energy-based DNNs can be used to jointly model input and output variables more richly, without the need to hand-specify dependency structures, leveraging DNNs’ impressive ability to learn correlation-capturing representations. Then I will introduce how my models can be trained using easily-provided human domain knowledge or other indirect supervision when labeled data is not available––making structured prediction available to a wider array of non-expert practitioners. Finally, I will discuss my future plans for energy-based DNNs in a variety of application domains, including image captioning and machine translation.

Pedram Rooshenas is a post-doctoral research fellow at the College of Information and Computer Sciences, University of Massachusetts, Amherst working with Prof. Andrew McCallum. He received his Ph.D. from the Department of Computer Science at the University of Oregon in 2017 with a dissertation on “Learning Tractable Graphical Models”. He is also the author of the LIBRA toolkit for learning and inference in graphical models. Pedram’s research interests are focused on the practical use of machine learning in complex problems, including innovations in graphical models, structured prediction, and deep learning.


Host: Pedro Szekely

POC: Alma Nava

« Return to Events