A novel method to improve transfer learning based on mahalanobis distance

作者：Yi, Chang'an^*; Zhen, Jiangjie; Li, Yang; Yi, Yang; Yin, Pengshuai; Min, Huaqing

来源：2017 IEEE International Conference on Robotics and Biomimetics, ROBIO 2017, Macau, China, 2017-12-05 To 2017-12-08.

DOI：10.1109/ROBIO.2017.8324758

• Classification accuracy
• Data classification
• Developmental robot
• K-nearest neighbors
• Mahalanobis distances
• Market information
• Metric learning
• Vision-based perception

Robots will be our partners in daily life, and they would not only manipulate objects but also read newspapers and analyze market information for us. In order to cope with the dynamic world, transfer learning becomes increasingly more important in action-based learning, vision-based perception and data-based analysis. In this work, our focus is data classification across different domains because data analysis is popular for any intelligent system especially for a developmental robot. Data attribute is often jointly determined by all of its dimensions. While most existing works do not consider the relationship among different dimensions of data, we use metric learning and k-nearest neighbor to improve the classification accuracy. Based on the source data that has already been labeled and a tiny proportion of the target data, a model is trained to predict the category of the rest target data. Experimental results indicate that our method performs better than traditional ones in the dataset.

• 出版日期2017