Minseo Yoon, Minseong Bae, Sojin Lee, Hyunwoo J. Kim · Weather LoRA adaptation

PROLA: Principal-Orthogonal Low-rank Adaptation

The paper introduces PROLA, a low-rank adaptation method that splits a fixed rank budget between the principal directions of pretrained weights and their orthogonal complement, and rescales rank-1 optimiser updates using gradient signals. It is applied to predictive spatiotemporal weather downscaling, where it reportedly outperforms representative low-rank adaptation baselines. The authors state the method also helps on diverse datasets and tasks beyond weather downscaling.

Open the source ↗

Similar Research & Safety releases

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.