🧑‍🔬 Eliza Rutherford

Eliza Rutherford's papers here concern large neural networks trained on vast amounts of text and images, within the areas of Language models & pretraining and Vision & language. Two questions run through the work. How should a limited training budget be divided between model size and data? And can a single model take on new visual tasks from a handful of examples rather than through task-specific fine-tuning?

The most cited of these papers is Flamingo: a Visual Language Model for Few-Shot Learning. It presents a family of models that read images and video together with text and produce free-form language. Shown only a few demonstrations in the prompt, they can adapt to new tasks. The second paper, Training Compute-Optimal Large Language Models, trains a large set of transformer language models across many sizes and data quantities. From these runs, it estimates in several independent ways how parameters and training tokens should grow together for a fixed amount of compute. The answer is that both should rise in step. By that measure, many large language models of the period were oversized and given too little training text.

Both papers date from 2022, so they also mark the most recent direction here: a move from studying how language models scale toward combining language with vision. The work was done in large collaborative teams. Arthur Mensch appears as a co-author on both papers. Other collaborators include Jean-Baptiste Alayrac, Sebastian Borgeaud and Lisa Anne Hendricks.

Eliza Rutherford has 2 papers in our library, in 2022, most of them on Language models & pretraining, Vision & language; often with Arthur Mensch, Aidan Clark, Jean-Baptiste Alayrac.