Pixel Language Models

Desmond Elliott

Assistant Professor

Jonas F. Lotz

Ph.D Fellow

(co-advised with Christian Igel)

Emanuele Bugliarello

Ph.D 2023. Now Research Scientist at Google

The overall goal of this line of research is to develop a new family of language models that can process any written language by rendering text as images, which allows the models to learn from the visual similarities between written languages. Realizing this goal includes creating and implementing tokenization-free multilingual language models, the collection and curation of visually diverse language data, the training of small-scale and large-scale models, developing techniques for effective model quantization and compression, and creating models that jointly process natural images and rendered text using a single interface. The project is currently funded by a grant from the Villum Foundation.