Tighter error bounds for representation learning and lifelong learning

点击翻译按钮获取中文摘要

The success of machine learning algorithms crucially hinges on which numerical features are used to represent data. An arguably large part of the success of human learning is that we not only learn representations to perform well on prediction tasks, but we also reuse these representations to more efficiently learn newly encountered but similar tasks. We are now in an era where "deep" machine learning methods actually learn useful representations of data automatically. These deep learning methods are becoming vital to many other fields: just a few successful applications include drug design, 3D rendering for computer graphics, and beating top-level human players in games such as Go. Despite these great empirical successes, our mathematical understanding for why these methods work well and how to best train them is lacking. Part I of my research is to develop new theory to understand how well a deep model will perform when it makes predictions about new data. My strategy to obtain better performance guarantees is to use theory that leverages specific properties of the actual data an algorithm sees and specific properties of the algorithm itself. By considering both of these important aspects of training, I expect to achieve success for deep learning models as well. Another important part of my analysis is to answer the following question: If a deep learning model is insensitive to certain types of transformations of its input, can it successfully be trained using less data as a result?One of the greatest advancements the field of machine learning can make is to shift from learning each new thing in isolation to reusing what has been learned in the past when learning new tasks. This continual transfer when learning an endless sequence of tasks is known as lifelong learning. While some research has begun in this important area, the mathematical theory for how well algorithms can perform lifelong learning is lacking, especially with regards to adaptive algorithms that transfer much more from their past experiences when they encounter tasks that are highly similar to past tasks. Part II of my research is to design algorithms for lifelong learning, including developing mathematical guarantees for these algorithms. I aim to answer questions like the following: If a learning agent encounters a series of slowly changing learning tasks, can we pool the data from previous tasks in order to learn new tasks using much less data? My strategy will be to adapt powerful ideas from sequential prediction, a vital yet previously untapped resource for work in lifelong learning.Both parts of this work can fundamentally advance the field of machine learning, which already is revolutionizing a number of sciences and industries. The students participating in this research will be excellently-equipped to fuel Canada's technological contributions at the global stage.