🧑‍🔬 Cade Gordon

Cade Gordon's papers here sit in the area of vision and language, specifically the large models that learn jointly from images and their accompanying text. The recurring concern is openness at scale. Training sets and models of this kind had mostly stayed inside a few industry labs, and the work asks what becomes possible when web-scale image-text data, training code and experiments are made public and reproducible.

The most cited of these works is LAION-5B, from 2022. It assembled a publicly released collection of billions of image-text pairs by gathering pictures and their alt-text from Common Crawl and downloading them in a distributed fashion. It then kept only pairs that an existing CLIP model judged sufficiently similar. To show that the collection is usable, the authors trained OpenCLIP models on a subset of it. A companion paper, Reproducible scaling laws for contrastive language-image learning, built on that resource. It used the open OpenCLIP code and public LAION subsets to measure how CLIP performance changes as the vision encoder, the amount of data and the number of training samples grow.

Both papers appeared in 2022, and the newer emphasis is on the second of these steps. Having an open dataset made it possible to study model behaviour systematically rather than only to release data, and the scaling study treats contrastive image-text pre-training as something any group can replicate and extend.

Gordon is affiliated with the University of California, Berkeley. The work is collaborative and closely tied to the LAION and OpenCLIP efforts. Recurring co-authors include Christoph Schuhmann, Mehdi Cherti, Romain Beaumont, Ross Wightman and Mitchell Wortsman. Further collaborators include Jenia Jitsev and Ludwig Schmidt.

Cade Gordon has 2 papers in our library, in 2022, most of them on Vision & language; often with Christoph Schuhmann, Mehdi Cherti, Mitchell Wortsman.