Motivated bythesuccess ofdiffusion models inhigh-fidelity and photorealistic image generation, generative data augmentation is an emerging application of diffusion models.

It has achieved great success on different vision-language downstream tasks suchasImage-TextRetrieval,Image-Captioning,andVisualQuestionAnswering.

Reasoning about the physical world requires models that are endowed with the right inductive biases to learn the underlying dynamics.

Standard FL approaches train a unique model for all clients [47, 32, 38, 29, 48].

In this setting, minibatch gradient estimates are exact (namely, exactly 0) on the set of global minimizers since data can be interpolated exactly.

Consider GofthelossL(w), meaning g 2G, L(w)= L(g w).