Enhancing Inter-Class Representation with a New Global Center Loss
Myeong K. Kang, In H. Lee, Eun Hee Lee, Sang Y. Baek, Sang Cheol Lee · 2017
The challenge of training convolutional neural networks (CNNs) is mapping a set of high dimensional low data to a lower dimensional feature point on the manifold. However, the dimensionality reduced feature from CNN always suffers from verifying a new unseen class, such as measuring a computably meaningful distance on the dimensionality distorted space. The mapped feature on the lower dimension is formed by interrelated with training dataset and measure of penalization of learning. These methods can boost performance by training a large dataset or applying extra penalizations, which help forming feature more discriminatively. The discriminatively learned feature has a balanced inter-class distance and a reduced intra-class variation. In this paper, we propose a new loss function so called global center loss to extract more meaningful distance on feature space. The method leads to a sufficient inter-class variation which helps forming feature more discriminatively. The insufficient enforcement of negative-log likelihood or local center loss can be complementarily enhanced by utilizing the proposed global center method which is a valuable joint supervisory signal.