Abstract by Taylor Archibald
An Autoregressive, Boundless Chirographer
Current off-line handwriting recognition models have not matched human-level performance in most tasks. Improvements in handwriting recognition are limited, in part, by the limited availability of labeled training data. If existing training data can be augmented with diverse, synthetically created handwritten data, handwriting recognition models can be improved directly. While traditional approaches to augmenting handwriting data typically involves perturbing existing handwriting samples, we propose the Autoregressive, Boundless Chirographer (ABC), an innovative approach which combines generated on-line handwriting samples with image-to-image translation to create novel, off-line handwriting samples to augment training data.