Integrating Supervised Laplacian Objective with CNN for Object Recognition.

Methods to improve object recognition accuracies of convolutional neural networks (CNNs) mainly focus on increasing model complexity and training samples, introducing training strategies, etc. Alternatively, in this paper, inspired by “manifolds untangling” mechanism from human visual cortex, we propose a novel and general method to improve object recognition accuracies of CNNs by embedding the proposed supervised Laplacian objective (SLO) into a high layer of the models during the training process. The SLO explicitly enforces the learned feature maps with a better within-manifold compactness and between-manifold margin, and it can be universally applied to different CNN models. Experiments with shallow and deep models on four benchmark datasets including CIFAR-10, CIFAR-100, SVHN and MNIST demonstrate that CNN models trained with the SLO achieve remarkable performance improvements compared to the corresponding baseline models.