The Golem —

An Artificial Intelligence Typography Experiment.

2021

Can AI give the Golem agency in the form of a Hebrew typeface it can communicate in?
Could a machine learning model be trained to generate realistic legible typefaces, in a language it cannot understand?

These questions set off the research into generative image tools, with the hypothesis that the machine would be able to generate images of typefaces that contained recognizable, yet new letterforms. These new generated typefaces would be visually based on all the typefaces given to the algorithm.

By using machine learning, instead of having the designer producing the content (the letterforms of the typeface), the designer is only responsible for the data input and duration of the experiment (how many training steps the model would have). The algorithm is then able to assess the dataset, take an action on it, compare the data against itself, determine likeness, and repeat the process if necessary. Once the desired level of similarity is achieved, the data can then be processed into an output - the generated typeface.

— The Approach

First step in setting up the machine learning model was to create the dataset. This dataset was constructed of 150 glyphs per typeface (hebrew letters, latin upper and lowercase letters, numerals, punctuation, currency and math signs) of all available hebrew typefaces on Google Fonts, Adobe Fonts and other free fonts, organised as individual image-per-glyph dataset.

A styleGAN2 model was picked to be trained in RunwayML on this dataset.

But first, what are GANs?

Generative Adversarial Networks (GANs) are one of the most interesting ideas in computer science today (the one you probably know from generating realistic looking human faces). Two models are trained simultaneously by an adversarial process. A generator learns to create images that look real, while a discriminator learns to tell real images apart from fakes.
During training, the generator progressively becomes better at creating images that look real, while the discriminator becomes better at telling them apart. The process reaches equilibrium when the discriminator can no longer distinguish real images from fakes.

The machine learning algorithm was trained using this dataset of images as reference so that it could then begin to recognize similarities between each glyph. A larger amount of training steps will result in an output that more closely resembles the original data.

These expermentations generated output showing a new typeface per frame—and an oscillation between different styles and kinds of typefaces, if you put one output frame after the other as a sequence.

Next step was to take these generated frames, randomise their order and upload them to Google’s Teachable Machine, there using an image classification model that was trained on the original labelled dataset of 150 images of each of the 150 glyphs, to select the first letterform for each glyph that reaches a threshold of a high accuracy rate.

After many trials and experimentation, it was apparent that as the training continued, the results began resembling designed letterforms to the point where it was hard to distinguish the generated letterforms from the trained data. Some characters took more training steps to create recognisable forms, but generally speaking each glyph was processed to a stage where the image classification model could recognise it after 2000 Training steps. Some letters such as ע (ein), ש (shin) and צ (tsadik) took substantially longer time to train. Other less complex glyphs were generated pretty quickly. Letting the machine have the ability to select which typeform would represent each glyph generated inconsistency in the form of the typeface, ranging from simple non serif letters to serif ones, and everything in between. This alphabet came out very crude, a ransom-letter-resembling assemblage of letterforms that may not be considered a typeface, had it been created solely by a designer.