Universal Hypernetworks: One Fixed Generator Predicts Weights for Any Target Architecture
arXiv·medium signal
Universal Hypernetwork (UHN) introduces a fixed-architecture generator that predicts weights from deterministic parameter, architecture, and task descriptors, completely decoupling the generator from target-network parameterization. Conventional hypernetworks require redesign and retraining when the target architecture changes — UHN eliminates this by using descriptor-based formulation. One trained UHN can serve arbitrary new architectures without retraining, enabling rapid prototyping of model variants.