Differentiable Programming for Computer Vision and Medical Image Analysis

深度学习已经成为一种强有力的方法，具有经验证据，可以为各种计算机视觉应用提供准确的学习模型。深度学习背后的成功似乎是需要大量标记图像和视频数据的监督学习。对于医学图像分析应用程序来说，需要大量带有专家创建的标签或标签的训练数据，这是一个特别具有挑战性的情况，因为训练数据的稀缺是很常见的。多年来，一些方法，如迁移学习已经出现，以缓解这个问题。然而，这些技术都不能充分解决缺乏训练数据的问题。拟议的研究计划将探索现有技术的替代方案，并将利用先验和领域知识以及深度学习的力量来解决标记训练数据的稀缺性。作为技术解决方案，该方案将依靠可微编程将传统的计算机视觉算法与深度学习方法相结合，其中先验知识可以包含在传统方法中。可微规划是指基于梯度下降的优化，它依赖于数据处理管道中使用的所有函数的可微性。然而，大多数传统的计算机视觉方法都包含不可微的函数或过程。因此，研究进一步提出利用旁路神经网络逼近不可微功能模块来克服这一困难。通过使用几个用例，提出的研究还表明，范围进一步扩大了训练数据的缺乏。例如，用于多目标跟踪的端到端检测跟踪系统可以纳入所提出的优化框架。该研究项目将有充足的机会为学生提供计算机视觉、医学图像分析、学习算法的理论和统计分析等方面的广泛培训，这些都符合加拿大对人工智能研究的承诺。提出的研究可以在计算机视觉和图像分析研究中创建一个新的和重要的学习算法家族。

Deep learning has emerged as a strong method with empirical evidence to deliver accurate learning models for various computer vision applications. The success behind deep learning seems to be supervised learning that requires lots and lots of tagged image and video data. The requirement to have lots of training data with expert-created labels or tags pose a particularly challenging situation for medical image analysis applications, where scarcity of training data is common. Over the years, some methods, such as transfer learning have emerged to mitigate this issue. However, none of these techniques can adequately address lack of training data. The proposed research program will explore an alternative to the exiting techniques and will make use of prior and domain knowledge along with the power of deep learning to address the scarcity of labeled training data. As a technical solution, the program will rely on differentiable programming to mix traditional computer vision algorithms with deep learning methods, where the prior knowledge can be included in the traditional methods. Differentiable programming refers to gradient descent based optimization that relies on the differentiability of all the functions used in a data processing pipeline. However, most traditional computer vision methods include functions or processes that are not differentiable. Thus, the research further proposes to overcome this difficulty by using bypass neural networks to approximate non-differentiable functional modules. Using several use cases, the proposed research also demonstrates that the scope extends further beyond the lack of training data. For example, an end-to-end detection-tracking system for multi-object tracking can be cast into the proposed optimization framework. The research program will have ample opportunity to provide broad training to students in computer vision, medical image analysis, theoretical and statistical analysis of learning algorithms that fit well into Canada's commitment to artificial intelligence research. The proposed research can create a new and significant family of learning algorithms in computer vision and image analysis research.