Combining knowledge with data for efficient and generalizable visual learning

doi:10.1016/j.patrec.2017.11.013

2019-06-01

PRE

Qiang Ji *

DOI：10.1016/j.patrec.2017.11.013

原文链接

原文求助

摘要

En 中文

Substantial progress has been made in the past decades in computer vision, in particular as a result of the application of deep learning methods. Despite these rapid developments, there still exist a significant gap between computer vision and human vision. One factor contributing to this gap is the data-driven and purely bottom-up nature of the existing visual learning methods and their inability to use prior knowledge. The data-driven and bottom-up approaches often cannot generalize well beyond the data that is used to train them. Parallel to data, there is usually prior knowledge in many domains that governs the target object, its context, and the computer vision tasks. Such knowledge, if utilized properly, can not only improve visual recognition performance but also reduce our dependence on data. To this goal, we propose to identify the related prior knowledge from different sources and to systematically encode them into visual learning tasks though joint bottom-up and top-down inference. Specifically, we first identify four types of prior knowledge, including permanent theoretical knowledge, circumstantial knowledge, subjective experiential knowledge, and data knowledge. We then demonstrate how permanent theoretical knowledge and circumstantial knowledge can be identified for different vision tasks and introduce methods to systematically represent and integrate them with the image data. Experiments on benchmark datasets show that by employing related prior knowledge, we can produce vision algorithms that are more data efficient, robust and generalizable and that are less dependent on training data. (C) 2017 Elsevier B.V. All rights reserved.

Keyword:

Computer vision

Machine learning

Object recognition