Deep Adaptive Networks for Visual Data Classification

This paper proposes a classifier called deep adaptive networks (DAN) based on deep belief networks (DBN) for visual data classification. First, we construct a directed deep belief nets by using a set of Restricted Boltzmann Machines (RBM) and a Gaussian RBM via greedy and layerwise unsupervised learning. Then, we refine the parameter space of the deep architecture to adapt the classification requirement by using global gradient-descent based supervised learning. An exponential loss function is utilized to maximize the separability of different classes. Moreover, we apply DAN to visual data classification task and observe an important fact that the learning ability of deep architecture is seriously underrated in real-world applications, especially when there are not enough labeled data. Experiments conducted on standard datasets of different types and different scales demonstrate that the proposed classifier outperforms the representative classification techniques and deep learning methods.