Task-driven dictionary learning based on convolutional neural network features

Modeling data as a linear combination of a few elements from a learned dictionary has been used extensively in the recent decade in many fields, such as machine learning and signal processing. The learning of the dictionary is usually performed in an unsupervised manner, which is most suitable for regression tasks. However, for other purposes, e.g. image classification, it is advantageous to learn a dictionary from the data in a supervised way. Such an approach has been referred to as task-driven dictionary learning. In this work, we integrate this approach with deep learning. We modify this strategy such that the dictionary is learned for features obtained by a convolutional neural network (CNN). The parameters of the CNN are learned simultaneously with the task-driven dictionary and with the classifier parameters.