【作者】 张智；

【导师】 唐降龙；

【作者基本信息】 哈尔滨工业大学 ， 计算机技术（专业学位）， 2018， 硕士

【摘要】 大部分机器学习算法都假设训练集和测试集处于同一特征空间并且具有一样的分布。然而在很多现实应用中,这种假设可能是不成立的。例如,假如要对某个域A内的数据进行分类任务,但该域内的数据量不够多,与此同时,在另一个相似的域B内拥有足够的数据训练模型,但是,域B内的数据有可能与域A内的数据处于不同的特征空间或者不服从同一分布。这时,如果成功进行知识迁移不仅可以极大地提升机器学习的性能而且还能减少人工标注数据的成本。最近这些年,迁移学习越来越吸引研究者的目光。深度学习是一种可以对数据的内在特征进行学习的方法,它通过模拟人脑构造多层神经网络对数据进行解释,例如图像,声音和文本等。由于Alex Net在2012年的Image Net大赛上获得冠军,深度学习在计算机视觉的研究领域获得了巨大成功,受到众多研究者的关注。但是深度学习需要大量的训练数据,获得这些训练数据将会耗费大量的人力物力,如果要在一个新的相似域内训练新的模型,丢弃这些大量的,不同分布下的训练数据将是非常浪费的。因此,把深度学习与迁移学习技术结合起来将能在提高模型精确度的同时节约训练成本。本文将主要介绍两种深度迁移学习图像分类方法,具体贡献包括以下两点:(1)提出了一种基于VGG13的深度适配网络。这种网络将两个域内的数据首先利用卷积层提取特征,然后将网络的某些层的输出映射到某个特征空间内,通过减少分类器在源域上的分类错误率和源域与目标域的数据在该特征空间内的距离达到在目标域提高分类正确率的目的。(2)提出一种基于胶囊网络的深度对抗网络。生成对抗网络的提出在深度学习域带来了一股新的潮流。深度网络的卷积层可以提取数据的特征,之后的全连接层对这些特征进行分类,深度对抗网络利用生成对抗网络的思想,在卷积层之后添加一个域分类器。假如域分类器能够对数据的域进行正确分类,那么只需要训练卷积层使其提取的特征无法被域分类器正确分类,就能得到域间不变的特征进行分类。胶囊网络是最近提出的一种卷积神经网络的改进,本文将提出一种利用对抗网络与胶囊网络结合进行迁移学习的方法。更多还原

【Abstract】 An important hypothesis for many machine learning and data mining algorithms is that training data and test data are in the same feature space and have the same distribution.However,in many real-world applications,this assumption may not be true.For example,if you want to classify data in a certain domain A,but the amount of data in the domain is not enough,at the same time,there is enough data in another similar domain B to train the model,but the data in domain B may be in a different feature space or disobey the same distribution with the data in domain A.At this time,successful knowledge transfer will greatly improve machine learning performance and reduce the cost of ma nually annotating data.In recent years,transfers learning has emerged as a new learning framework.Deep learning is a method of characterizing learning based on data in machine learning.It is a new field in machine learning.Its motivation lies in building and simulating the neural network of the human brain to analyze and learn.It imitates the mechanism of the human brain.To interpret data such as images,sounds,and text.Since Alex Net won the championship in the Image Net competition in 2012,deep learning has achieved great success in the field of computer vision research and has attracted the attention of many researchers.However,deep learning requires a large amount of training data.Obtaining these training data will consume a lot of manpower an d material resources.If a new model is to be trained in a new similar field,discarding these large amounts of training data under different distributions will be very wasteful.of.Therefore,the combination of deep learning and transfer learning techniq ues will improve training accuracy while improving model accuracy.This paper will mainly introduce twodeep transfer learning classification methods,and compare their advantages and disadvantages.Specific mo dels include the following two:(1)Deep adaptation network.This kind of network will first process the data in two domains through deep network processing,and then map the output of some layers of the network into a certain feature space.By reducing the classification error rate of the classifier i n the source domain and the distance between the data in the source domain and the target domain in the feature space,the purpose of improving the classification per formance in the target domain can be achieved.(2)Deep adversarial network.The Generative adversarial network has brought a new wave in the field of deep learning.The con volutional layer of the deep network can extract the feature of the data,and then the full-connected layer classifies these features,deep adversarial network utilizes this idea,and adds a domain classifier after the convolutional layer.Assuming that the domain classifier can correctly classify the data domain,then only the convolution layer needs to be trained so that the extracted features cannot be correctly classifie d by the domain classifier,and the domain-invariant features between the domains can be obtained.Using this kind of domain-invariant features can also achieve the purpose of transfer learning.更多还原