🏢 Qualcomm AI Research

On-device training with fixed-point forward gradients enables efficient model personalization on resource-constrained edge devices, overcoming backpropagation’s memory limitations.

Two novel sampling strategies for spectral model sharding in federated learning minimize approximation error and create unbiased estimators, improving performance on various datasets.

Hollowed Net efficiently personalizes text-to-image diffusion models on-device by temporarily removing deep U-Net layers during training, drastically reducing memory usage without sacrificing performa…

FouRA: a novel low-rank adaptation method improves text-to-image generation by learning projections in the Fourier domain and using an adaptive rank selection strategy, addressing LoRA’s limitations o…

This paper uses information theory to improve conformal prediction, proving new ways to bound uncertainty and creating better training methods and side-information incorporation.