Abstract

n the last few years, natural language processing and computer vision have experienced a fundamental shift in the way these fields use machine learning. Rather than training neural networks from a randomly initialized set of parameters, researchers have often found superior performance can be achieved by fine-tuning a general pre-trained “foundation model” trained on vast amounts of diverse data – perhaps because this model comes with better “priors” than an untrained network. Polymathic AI1 is a new research collaboration that aims to usher in the same shift in machine learning for scientific datasets. In this talk I will present the motivations behind the collaboration and describe the findings of our three new papers in this space, which examine: better numerical encodings for large language models2, contrastive embeddings for multi-modal scientific data3, and building machine learning models that learn from multiple types of physics4.

1 https://polymathic-ai.org/ 2https://arxiv.org/abs/2310.02989 3https://arxiv.org/abs/2310.03024 4https://arxiv.org/abs/2310.02994