Sequence Generation with a Physiologically Plausible Model of Handwriting and Recurrent Mixture Density Networks

The purpose of this study is to explore the feasibility and potential benefits of using a physiological plausible model of handwriting as a feature representation for sequence generation with recurrent mixture density networks. We build on recent results in handwriting prediction developed by Graves (2013), and we focus on generating sequences that possess the statistical and dynamic qualities of handwriting and calligraphic art forms. Rather than model raw sequence data, we first preprocess and reconstruct the input training data with a concise representation given by a motor plan (in the form of a coarse sequence of ‘ballistic’ targets) and corresponding dynamic parameters (which define the velocity and curvature of the pen-tip trajectory). This representation provides a number of advantages, such as enabling the system to learn from very few examples by introducing artificial variability in the training data, and mixing of visual and dynamic qualities learned from different datasets.