Abstract:

Transfer learning leverages the rich data in the source domain to provide support for building accurate models in the target domain. Feature transfer learning is a kind of widely studied technology in transfer learning, but the existing feature transfer methods are facing with the following problems. Firstly, some existing methods can only implement linear feature transfer learning, so the ability of these methods to transfer learning is limited. Secondly, other kinds of methods can achieve nonlinear feature transfer learning, while it is often necessary to introduce strategies such as kernel techniques, which makes the process of feature transfer difficult to understand. In view of these problems, this paper introduces fuzzy reasoning technology and proposes a feature transfer method based on uncertain reasoning rules. The proposed method uses fuzzy inference system to realize feature transfer and uses manifold regularization to avoid information loss during feature transfer learning. Because the fuzzy system has good nonlinear modeling ability and good interpretation, the proposed method has good nonlinear feature transfer ability and is easy to understand the obtained new features. A large number of experiments have proven that the proposed method can be significantly better than the existing methods in the cross-domain image classification problem.

Key words: feature transfer, nonlinear model, kernel technique, fuzzy system, interpretability

摘要：

迁移学习利用源域中丰富的数据来为目标域构建精确的模型提供辅助和支持。特征迁移学习是迁移学习中被广泛研究的一类技术，但是现有的特征迁移方法面临着如下的问题：一些已有的方法仅能实现线性的特征迁移学习，因此这些方法迁移学习的能力有限。另一类方法虽然能实现非线性特征迁移学习，但往往需要引进核技巧等策略，这使得特征迁移的过程难以理解。针对此，引入模糊推理技术，提出基于不确定推理规则的特征迁移方法。该方法基于模糊推理系统来实现特征迁移，并利用流形正则化技术来避免特征迁移过程中的信息损失。由于模糊系统具有很好的非线性建模能力以及基于规则的良好的解释性，因此提出的方法具有良好的非线性特征迁移能力，并易于对新特征进行理解。大量实验证明，该算法在跨域图像分类问题上可以明显优于已有的多种方法。

关键词: 特征迁移, 非线性模型, 核技巧, 模糊系统, 解释性