Conference on Cognitive Computational Neuroscience

Handwriting has been utilized to diagnose motor-related neurodegenerative diseases and developmental standardization because the skill involves the dynamic interactions of sensation, motor control, and conceptualization. Recent deep neural networks successfully mimic realistically shaped handwriting, which implies they are propitious models for biological motor control. However, relevant studies have focused on shapes and not suggested methods that ensure the realistic dynamics of motor control. Therefore, we propose these methods combined into a single framework to develop a controller that generates motor commands of handwriting endpoints and to evaluate whether the model generates as spatiotemporally realistic samples as a human does. Our evaluation pipeline could confirm that this neural network controller generated realistic shapes and motor dynamics. With this certainty to achieve realistic dynamics, we have investigated the model’s implications for human handwriting. We show that models learned handwriting in three stages, which resemble the characteristics discovered in the human learning stages: learning to draw character shapes, then reducing variability, and finally keeping certain variability. Finally, we exhibit that the models deprived of the previous movement feedback generated fluctuating handwriting with increasing speed inversions yet recognizable shapes, which was also found in the handwriting of a patient lacking proprioceptive feedback.