Unified Classification and Generation Networks for Co-Creative Systems

This paper reports on a new deep machine learning architecture to classify and generate input for co-creative systems. Our approach combines the generational strengths of Variational Autoencoders with the image sharpness typically associated with Generative Adversarial Networks, thereby enabling a generative deep learning architecture for training co-creative agents called the Auxiliary Classifier Variational Autoencoder (AC-VAE). We report the experimental results of our network’s classification accuracy and generational loss on the MNIST numerical image dataset and TU-Berlin sketch data set. Results indicate our technique is effective for classifying and generating sketched object images, with larger sizes. We also describe how our network is particularly useful for co-creative agents since it can generate diverse concepts, as well as transform and morph user generated sketches while maintaining their concept identity.