Deep learning for sequence pattern recognition

Deep Learning is a superb way to solve remote sensing related problems, which mainly cover four perspectives: image processing, pixel-based classification, target recognition and scene understanding. In this paper, we focus on target recognition by building deep learning models, and our target is sequence pattern. Accurate prediction of sequence pattern would help identify significant characters from text sequence. Despite considerable advances in using machine learning techniques for sequence pattern recognition problem, its efficiency is still limited because of its involving extensive manual feature engineering in the process of features extraction from raw sequences. Thus, we apply a deep learning approach in sequence pattern recognition problem. The sequences of the datasets we used are self-generated genomic format sequences, and each dataset is generated based on a kind of pattern. We then investigate and construct various deep neural network models (such as convolutional networks, recurrent networks and a hybrid of convolutional and recurrent networks). The one-hot encoding method that preserves the vital position information of each character is presented to represent sequences as inputs to the models. The sequence patterns are then extracted from the input and output the probabilities of the existence of sequence patterns. Experimental results demonstrate that the deep learning approaches can achieve high accuracy and high precision in sequence pattern recognition. In addition, a saliency-map-based method is applied to visualize the learned sequence patterns. In view of the simulation results, we believe that we can find an appropriate deep learning model for a certain sequence sensing problem.