Expressive data augmentation in deep learning

点击翻译按钮获取中文摘要

Deep learning is a subfield of machine learning, a type of artificial intelligence, whose goal is to automatically learn how to solve problems using data. For example, a typical task in deep learning is called "image classification", and consists of learning how to categorize images into different categories (e.g. "cat", "dog", "human") when given a dataset of images labeled with their corresponding category. For a human, this task is easy, but since computers can only "see" an image as a bunch of ones and zeros, it is hard to encode how to solve the problem into an algorithm, a set of mechanical instructions that a computer can follow. In recent years, deep learning has enabled new applications where designing such algorithms is difficult, making advances on tasks involving images, audio, and language, among a variety of others. One key limitation of deep learning is that it typically requires a large amount of data in order to work well. In image classification, for example, several thousand to several million labeled images are required for reasonable performance, a prohibitive cost for most applications. To help compensate, it is common to artificially generate new data from existing data, a process known as "data augmentation". A basic example of this is to randomly make slight alterations to an image's brightness while maintaining its label - an image of a cat is still an image of a cat, even if the brightness is changed by a small amount. Data augmentation has the effect of expanding the effective size of the dataset used to learn algorithms without requiring the costly collection of new data. Despite its large utility, a number of challenges exist when using data augmentation, which my research intends to solve. When applying it to a new problem, for example, one needs to define its basic operations (e.g. the random brightness change itself) and decide on their precise strengths, which may be costly. My research will define augmentation operations automatically by learning operations that vary the precise appearance of images while preserving their desired labels. Then, to tune the strength of each operation, my research will investigate the learned behavior of algorithms trained without data augmentation; if algorithm output varies greatly with respect to a particular operation, then it is likely that using strong amounts of it as data augmentation will improve an algorithm's robustness to it. If successful, my research will allow for the automatic creation of expressive data augmentation policies, substantially reducing the amount of data required to unlock new applications of deep learning throughout both research and industry in Canada. One particularly exciting application of this is in medicine, since the data available for most medical tasks is limited. Ideally, the development of both improved and novel diagnostics may be possible, advancing Canadian medical research and eventually the health of the Canadian public as a whole.