A kernel-based weight decorrelation for regularizing CNNs

Abstract

Abstract Recent years has witnessed the success of convolutional neural networks (CNNs) in many machine learning and pattern recognition applications, especially in image recognition. However, due to the increasing model complexity, the parameter redundancy problem arises, and greatly degrades the performance of CNNs. To alleviate this problem, various regularization techniques, such as Dropout, have been proposed and proved their effectiveness. In this paper, we propose a novel adaptive kernel-based weight decorrelation (AKWD) framework, in order to regularize CNNs for better generalization. Different from existing works, the correlation between paring weights is measured by the cosine distance defined in RKHS associated with a specific kernel. The case with the well-known Gaussian kernel is investigated in detail, where the bandwidth parameter is adaptively estimated. By regularizing CNN models of different capacities using AKWD, better performance is achieved on several benchmark databases for both object classification and face verification tasks. In particular, when Dropout or BatchNorm is present, even higher improvements are obtained using the proposed AKWD, that demonstrates a good compatibility of the proposed regularizer with other regularization techniques.