We have started this project by demonstrating the advantages of visual analytics techniques to understand the deep learning process and to select features for more effective classification. In particular, we have developed ways to visualize and understand the neural activity during back propagation and to identify which neurons are specialized to which categories of the classification problem. We have also assessed techniques to find effective Convolution Networks (ConvNets) for feature learning (characterization) under a limited number of supervised samples, as an important topic to reduce human effort and time in label supervision. We are currently investigating how visual analytics can be useful to improve feature learning based on ConvNets, Visual Dictionaries, Autoencoders, by guiding the intervention of specialists in the process.