Abstract
Training of deep neural networks heavily depends on the data distribution. In particular, the networks easily suffer from class imbalance. The trained networks would recognize the frequent classes better than the infrequent classes. To resolve this problem, existing approaches typically propose novel loss functions to obtain better feature embedding. In this paper, we argue that drawing a better decision boundary is as important as learning better features. Inspired by observations, we investigate how the class imbalance affects the decision boundary and deteriorates the performance. We also investigate the feature distributional discrepancy between training and test time. As a result, we propose a novel, yet simple method for class imbalanced learning. Despite its simplicity, our method shows outstanding performance. In particular, the experimental results show that we can significantly improve the network by scaling the weight vectors, even without additional training process.
Highlights
In deep neural networks (DNN), imbalanced data distribution is critical as these networks learn directly from the data distribution
We provide an in-depth analysis based on observation and propose a simple yet powerful method for class imbalanced learning
We show that the bias in the decision boundary is closely related to the norm of each weight vector
Summary
One important observation is that minimizing the empirical loss while using a conventional training framework results in decision boundaries that allocate a larger volume of the feature space to more frequent classes. Kim: Adjusting Decision Boundary for Class Imbalanced Learning follows the sample frequency; more samples form a larger cluster. We achieved better performance than the existing methods without using any additional training process This suggests that we can obtain a feature extractor of fine quality by minimizing empirical loss and that the problems with class imbalanced learning mainly depend on how the appropriate decision boundary is drawn. We show that we can adjust the decision boundary by controlling the norm of the weight vector Another observation of ours shows positive correlation between the sample frequency and generalization. It implies that how the decision boundary is drawn is as important problem as how the feature representation is trained
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
