WMJi
commented
6 years ago Open WMJi opened 6 years ago
WMJi
commented
6 years ago WMJi
commented
6 years ago 理解《A Survey on Transfer Learning》 Pan S J, Yang Q. A Survey on Transfer Learning[J]. IEEE Transactions on Knowledge & Data Engineering, 2010, 22(10):1345-1359.
Abstract—A major assumption in many machinelearning and data mining algorithms is that the training and future data mustbe in the same feature space and have the same distribution. However, in manyreal-world applications, this assumption may not hold. For example, wesometimes have a classification task in one domain of interest, but we onlyhave sufficient training data in another domain of interest, where the latterdata may be in a different feature space or follow a different data distribution.In such cases, knowledge transfer, if done successfully, would greatly improvethe performance of learning by avoiding much expensive data-labeling efforts.In recent years, transfer learning has emerged as a new learning framework toaddress this problem. This survey focuses on categorizing and reviewing thecurrent progress on transfer learning for classification, regression, andclustering problems. In this survey, we discuss the relationship betweentransfer learning and other related machine learning techniques such as domainadaptation, multitask learning and sample selection bias, as wellas covariate shift. We also explore some potential future issues in transferlearning research.
传统机器学习的领域假设训练数据和测试数据属于相同的特征空间并在同一分布上。然而，现实应用中这种假设往往得不到满足。例如，我们对目标领域的分类问题感兴趣，却只有源领域的训练数据。但源领域数据与目标领域数据要么不在同一特征空间，要么不满足相同的数据分布，例如需要进行的文本分类语言是西班牙语，但只提供了葡萄牙语的文本。在某些情况下成功地进行知识迁移能够很大程度上提高学习的性能，也同时降低了标记目标领域数据带来的大量时间和人力成本。近年来，迁移学习已经成为一种解决知识迁移问题的新型学习框架。这篇论文讨论了使用迁移学习进行分类、回归以及聚类的一般过程，也讨论了迁移学习和其他相关的机器学习技术之间的关系，如领域适应性、多任务学习、样本选择和变量转换。
这篇综述十分详细地介绍了各类迁移学习方法，进行分类并介绍了各类方法的代表性文章。
迁移学习:将所有的图片交给 VGG16，利用 VGG16 的深度网络结构中的五轮卷积网络层和池化层，对每张图片得到一个 4096 维的特征向量，然后我们直接用这个特征向量替代原来的图片，再加若干层全连接的神经网络，在花朵特征向量集和花朵标签集上进行训练。本质上，我们是将 VGG16 作为一个图片特征提取器，然后在此基础上再进行一次普通的神经网络学习，这样就将原先的 244 224 3 维度的数据转化为了 4096 维的，而每一维度的信息量大大提高，从而大大降低了计算资源的消耗，实现了把学习物体识别中得到的知识应用到特殊的花朵分类问题上。