Deep Feature Representation via Multiple Stack Auto-Encoders

Recently, deep architectures, such as stack auto-encoders (SAEs), have been used to learn features from the unlabeled data. However, it is difficult to get the multi-level visual information from the traditional deep architectures (such as SAEs). In this paper, a feature representation method which concatenates Multiple Different Stack Auto-Encoders (MDSAEs) is presented. The proposed method tries to imitate the human visual cortex to recognize the objects from different views. The output of the last hidden layer for each SAE can be regarded as a kind of feature. Several kinds of features are concatenated together to form a final representation according to their weights (The output of deep architectures are assigned a high weight, and vice versa). From this way, the hierarchical structure of the human brain cortex can be simulated. Experimental results on datasets MNIST and CIRFA10 for classification have demonstrated the superior performance.