While these advances use different algorithmic techniques, e.g., contrastive learning, diffusion, or auto-regressive modeling, they all rest on a common foundation: large datasets containing paired image-text examples.

We find that the performance of the pre-training data varies substantially across distribution shifts, with no single data source dominating.

To mitigate the negative effect of data heterogeneity (non-IIDness), two common approaches are clustering and personalization.

Optimal Transport distance to predict the model's performance for any data mixture

In the United States, health insurance is a critical component of individuals' ability to access health care.