Tag
提出 SHINE，一种从稀疏历史搜索日志与零成本代理出发的跨任务神经架构搜索策略，通过组合各源任务的冠军架构并联合训练成对排序器，显著缓解单一知识源在目标任务上的严重失败，同时保持有竞争力的平均性能。
RiPPLE introduces a method for ranking neural architectures by training a small set of anchors to early stages, extrapolating learning curves, and propagating labels, achieving efficient and accurate performance prediction across search spaces.