Learning Discriminative Features using Encoder-Decoder type Deep Neural Nets

Researchers have solved pattern recognition problems (to varying degrees of success) like face detection [5], gender classification [13], human expression recognition [14], object learning [1], unsupervised learning of new tasks [8] and also have studied complex neuronal properties of higher cortical areas [9]. However, most of the above techniques did not require automatic feature extraction as a pre-processing step to pattern classification. In contrast to the above, there exist many practical applications characterized by high dimensionality of data (such as speech recognition, remote sensing, e.t.c), where finding sufficient labeled examples might not be affordable or feasible. At the same time there may be lot of unlabeled data available easily. Unsupervised feature learning techniques, like the Autoencoder ([7], [16], [3] and [20]), try to capture the essential structure underlying the high-dimensional input data by converting them into lower dimensional data without losing information.