Visual Recognition Beyond Supervised Learning

在过去的几年里，各种视觉识别任务的性能都有了显著的提高，如图像分类、目标检测、语义分割等。大部分成功是通过使用带有人类标注的大规模数据集来学习深层神经网络模型的暴力-完全监督方法实现的。然而，这种蛮力方法有几个限制，阻碍了视觉识别技术在许多现实世界的应用中的广泛采用。学习这些模型需要获得大规模的标签数据，而这些数据往往很难收集。标准的监督学习算法通常不考虑以前学习到的知识，对于每个新任务都必须从头开始学习。已学习的系统通常不能很好地适用于新的场景。我的研究计划的长期目标是开发算法，使学习视觉识别系统变得容易得多。我们希望能够在没有太多人工监督的情况下从小数据中学习视觉识别系统，通过利用以前学到的知识快速学习新任务，并能够积极探索和适应新环境。换句话说，目标是超越监督学习，转向建立基本上是自学的视觉识别系统。为了实现这一长期目标，我提出了几个短期目标，我的目标是在该提案的5年时间框架内实现：(1)开发视觉识别中知识转移的新技术；(2)开发弱监督的视觉识别算法；(3)将视觉识别和主动探索相结合。该项目将产生16个HQP培训在不同的水平(博士，硕士，理科)。这些HQP将接受计算机视觉、机器学习(特别是深度学习)和软件开发方面的研究培训。加拿大对这些技能的需求很高。离开这个研究项目的学生将很容易在学术界或行业找到工作。拟议的研究将产生在基础科学研究和实际应用方面都具有重要影响的新知识。这项研究解决了计算机视觉和机器学习中的一些基本研究问题，如小数据学习、迁移学习、领域自适应、弱监督学习、主动视觉等。同时，这项研究对视觉识别在现实世界中的应用具有巨大的潜力。

In the past few years, there has been significant improvement in the performance of various visual recognition tasks, such as image classification, object detection, semantic segmentation, etc. Most of the success is achieved by a "brute-force" fully supervised approach that learns deep neural network models using large-scale datasets with human annotations. However, this brute-force approach has several limitations that prevent the wide adoption of visual recognition technologies in many real-world applications. Learning these models requires access to large-scale labeled data which are often difficult to collect. Standard supervised learning algorithms usually do not take into account of previously learned knowledge and have to learn from scratch for every new task. The learned systems often do not generalize well to new scenarios. The long-term goal of my research program is to develop algorithms that make it far easier to learn visual recognition systems. We would like to be able to learn visual recognition systems from small data without too much human supervision, to learn new tasks quickly by leveraging previously learned knowledge, and to be able to actively explore and adapt to new environments. In other words, the goal is to move beyond supervised learning and move towards building visual recognition systems that are largely self-taught. Towards realizing this long-term goal, I propose several short-term objectives that I aim to achieve during the 5-year time frame of the proposal: (1) developing new techniques for knowledge transfer in visual recognition; (2) developing algorithms for visual recognition with weak supervision; (3) combining visual recognition and active exploration. This project will produce 16 HQP trained at different levels (PhD, MSc, BSc). These HQP will receive research training in computer vision, machine learning (especially deep learning), and software development. These skills are in high demand in Canada. Students leaving this research program will easily find employment in academia or industry. The proposed research will produce new knowledge that has important impact in terms of both fundamental scientific research and practical applications. The proposed research addresses some fundamental research questions in computer vision and machine learning, such as learning from small data, transfer learning, domain adaptation, weakly-supervised learning, active vision, etc. At the same time, this research has tremendous potential to impact how visual recognition is deployed in real-world applications.