Tech uses artificial intelligence to generate handwriting

In the context: Researchers are turning the creative world on its head by using artificial intelligence and machine learning algorithms to turn many tasks into semi-autonomous processes. Nothing is immune to generative AI anymore, not even the illegible handwriting of a local doctor.

Years before OpenAI and other organizations began experimenting with AI to easily generate text, speech, illustrations, malware, and video, machine learning researcher Sean Vasquez studied a 2013 article by Alex Graves of Google DeepMind to conduct experiments on handwriting synthesis.

Vasquez in the archive his code is on GitHub along with his web demo. The experiment is available at Calligrapher.aiwhich Hacker News recently reopened. Handwriting synthesis in uses a generative method based on a recurrent neural network (RNN).

RNN is a class of artificial neural networks where connections between nodes can create a loop, allowing the output of some nodes to influence subsequent inputs for those same nodes. Recurrent neural networks can exhibit temporal dynamic behavior, which makes them especially useful in tasks such as handwriting or speech recognition. Like any other neural network, Vasquez trained on a moderately large dataset of calligraphy samples, primarily on IAM’s online handwriting database.

The IAM-On database contains “blackboard handwritten English forms”, with samples from 221 different “writers” and over 1,700 received forms. The database includes 13,049 isolated and marked text strings in online and offline formats, totaling 86,272 samples from a dictionary of 11,059 words. can generate variable handwriting in 9 different styles, and users can change speed, legibility and stroke width with sliders for further customization. Unlike traditional font types designed to mimic handwriting, each swatch created by must be unique, even if the writing style is the same. Users can download the final result as a vector SVG file.

According to Vasquez, the readability slider uses a technique known as “selective distribution temperature adjustment” to change handwriting variation. The output comes from a “probability distribution” and increasing legibility “effectively concentrates the probability density around more likely outcomes”.

Being just a demo, is limited in features despite its ability to create believable handwriting patterns. Also, Vasquez only taught basic RNN on English samples, so the website is not particularly good at reproducing accents commonly used in other languages.

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button