Deep Learning and Representation Learning for Sequential Data

点击翻译按钮获取中文摘要

The pervasiveness of computing has resulted in the production and storage of more data than ever before. Machine learning seeks to transform this deluge of data into intelligent systems that identify patterns and make decisions. It has revolutionized fields as diverse as computer vision, computational neuroscience, biology and the social sciences. But when faced with data that is increasingly complex, how does a machine know which parts are relevant? How does it structure the millions of components into organized units on which it can base decisions?Recent developments in machine learning, known as "Deep Learning", have answered these questions by showing how increasingly abstract layers of features can extract informative data representations without human guidance. But to-date, the field has focused on static as opposed to dynamic data: particularly images in the context of visual reasoning. Much of the real-world data we encounter in machine learning applications, however, has temporal dependencies that often extend over large time scales. Examples are human or robot motion, climate data, audio (e.g. music or speech), and finance. Modeling sequences is challenging: even the most sophisticated techniques fail to learn long-term structure. Sequences increase computational requirements as we attempt to untangle more complex explanatory factors underlying the data.I propose a research program that confronts these challenges through the discovery of algorithms and architectures that learn representations from sequences. I also aim to widen the adoption and increase the relevance of Deep Learning outside of academia through interdisciplinary collaborations that will impact fields such as biology, entertainment, and finance. I will train high-quality personnel who can satiate industry's growing demand for data scientists.