TITLE: Scalability and Methods of Deep Learning from an Optimization Viewpoint
Kenji Kawaguchi is a Ph.D. candidate at Massachusetts Institute of Technology (MIT). His research interests span deep learning theory, machine learning theory, artificial intelligence, nonconvex optimization, and Bayesian optimization. His research aims to advance deep learning theory in order to make our society safer by enabling deep learning applications for safety critical systems, and also in order to understand how deep learning can scale to large problems in the future such as data science for healthcare, automatic transportation, and robotics. Using theoretical insights, he also develops AI methods for applications.
Deep learning is becoming more vital to our daily lives and our scientific endeavors, with applications in scientific computing, data analysis for high-energy physics, medical diagnosis, object detection, speech recognition, and more. As the amount of available data and computational power are increasing, we are facing a rise in the demand for deep learning and observing its great empirical success today. However, can the success of deep learning continue to scale for larger and larger problems that we will face in the future? What deep learning methods can we use for new problems of different scales in various applications? The empirical success today does not completely answer these questions because a large problem today is a small problem tomorrow and an empirical success of a method on one problem does not imply its success on a new problem. To answer these questions, I will present several recent results in deep learning theory. To motivate the discussion, I will also briefly touch on a few real-world applications of deep learning.