Cooperative auto-classifier networks for boosting discriminant capacity

Abstract

• A cooperative game between classifier and generator networks . • Auto-classifier: exploiting dimensionality reduction to enhance classifier capacity. • Auto-classifier has a better regularization error than a standalone classifier (CNN). • Auto-classifier outperforms MNIST and SVHN benchmark. Although the training accuracy of deep learning using a deep structure is high, the depth of the deep-learning structure is directly proportional to the generalization error . To address this issue, we propose the auto-classifier, a novel classifier that automatically exploits dimensionality reduction. The proposed classifier contains both classifier and generator networks. It is dedicated to generating separable outcomes based on the label and implicitly capturing the latent variable; simultaneously, the generator network must be able to reconstruct the original data based on the given latent variable. We introduce a cooperative learning mechanism with a new loss function that enables the classifier and generator networks to cooperate to achieve the aforementioned objectives. Extensive experiments were conducted using benchmark datasets. The results revealed that the accuracy of the proposed classifier, without any data augmentation, distortion, or pretraining mechanism, was very competitive with the existing state-of-the-art benchmark datasets.