【计算机科学】【2018.08】在深度学习领域中推进分割和无监督学习-白红宇

【计算机科学】【2018.08】在深度学习领域中推进分割和无监督学习

阅读量：244 次

发布时间：2019-02-28

本文共 2786 字，大约阅读时间需要 9 分钟。

在这里插入图片描述

本文为挪威北极大学（作者：Michael Kampffmeyer）的博士论文，共93页。

由于基于深度学习的模型给各种任务带来了巨大的改进，近年来得到了大量的关注。然而，这些改进在很大程度上是在有标签的监督设置中实现的，并且最初集中于传统的计算机视觉任务，例如视觉对象识别。考虑大尺寸和多模态图像的特定应用领域，以及标记训练数据难以获得的应用，反而受到较少的关注。本文旨在从两个方面填补这些空白。首先，我们提出了专门针对遥感和医学成像应用的分割方法。其次，受医学影像等高影响领域缺乏标记数据的启发，提出了四种无监督的深度学习任务：领域适应、聚类、表征学习和零镜头学习。分割的工作解决了类别不平衡、缺失数据模式和遥感不确定性建模的挑战。基于像素连通性的思想，我们进一步提出了一种新的显著性分割方法，这是一种常见的预处理任务。我们将该问题描述为连通性预测问题，可以在保持模型简单的同时获得良好的性能。最后，结合我们在分割和无监督深度学习方面的工作，提出了一种在医学领域的分割环境中无监督域自适应方法。除了无监督的领域自适应外，我们还提出了一种新的基于核方法思想和信息理论学习的聚类方法，取得了很好的效果。基于我们的直觉，有意义的表示应该包含数据点之间的相似性，我们进一步提出了一个核心化的自动编码器。最后，我们提出了基于改进图卷积神经网络知识传播的零镜头学习任务，在21K类ImageNet数据集上实现了最先进的性能。

Due to the large improvements that deeplearning based models have brought to a variety of tasks, they have in recentyears received large amounts of attention. However, these improvements are to alarge extent achieved in supervised settings, where labels are available, andinitially focused on traditional computer vision tasks such as visual objectrecognition. Specific application domains that consider images of large sizeand multi-modal images, as well as applications where labeled training data ischallenging to obtain, has instead received less attention. This thesis aims tofill these gaps from two overall perspectives. First, we advance segmentationapproaches specifically targeted towards the applications of remote sensing andmedical imaging. Second, inspired by the lack of labeled data in manyhigh-impact domains, such as medical imaging, we advance four unsupervised deeplearning tasks: domain adaptation, clustering, representation learning, andzero-shot learning. The works on segmentation address the challenges ofclass-imbalance, missing data-modalities and the modeling of uncertainty inremote sensing. Founded on the idea of pixel-connectivity, we further propose anovel approach to saliency segmentation, a common pre-processing task. Weillustrate that phrasing the problem as a connectivity prediction problem,allows us to achieve good performance while keeping the model simple. Finally,connecting our work on segmentation and unsupervised deep learning, we proposean approach to unsupervised domain adaptation in a segmentation setting in themedical domain. Besides unsupervised domain adaptation, we further propose anovel approach to clustering based on integrating ideas from kernel methods andinformation theoretic learning achieving promising results. Based on ourintuition that meaningful representations should incorporate similaritiesbetween data points, we further propose a kernelized autoencoder. Finally, weaddress the task of zero-shot learning based on improving knowledge propagationin graph convolutional neural networks, achieving state-of-the-art performanceon the 21K class ImageNet dataset.