DiffusionNAG: Conditional Neural Architecture Generation with Diffusion Models
Core Concepts
DiffusionNAG proposes a paradigm shift in Neural Architecture Generation by leveraging diffusion models to efficiently generate task-optimal architectures guided by property predictors.
Abstract
Existing NAS methods face challenges of high search costs and inefficiencies.
DiffusionNAG introduces a conditional NAG framework based on diffusion models.
The model generates architectures that meet specified conditions efficiently.
Extensive experiments validate the effectiveness of DiffusionNAG in Transferable NAS and BO-based NAS scenarios.
DiffusionNAG
Stats
DiffusionNAG achieves speedups of up to 35× on Transferable NAS benchmarks.
Quotes
"DiffusionNAG significantly outperforms existing NAS methods in such experiments."