Aviv Navon

GradMetaNet is a neural architecture that efficiently processes gradients of other networks by exploiting their symmetries and rank-1 decomposition structure, enabling better learned optimizers, model editing, and loss curvature estimation