arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2108.09671cs.CVcs.LG
收藏

Pi-NAS:通过减少超网训练一致性偏移来改进神经结构搜索

Pi-NAS: Improving Neural Architecture Search by Reducing Supernet Training Consistency Shift

Sun Yat-sen University(中山大学);DarkMatter AI Research(DarkMatter AI 研究院);Monash University(莫纳什大学);University of Oxford(牛津大学)

Jiefeng Peng, Jiqi Zhang, Changlin Li, Guangrun Wang, Xiaodan Liang, Liang Lin

摘要:

最近提出的神经架构搜索(NAS)方法在超网中共同训练数十亿个架构,并使用从超网分离的网络权重来估计它们的潜在准确性。然而,网络体系结构的预测精度和实际性能之间的排序关系是不正确的,这导致了现有的NAS方法的困境。我们将这种排名相关性问题归因于超网训练的一致性偏移,包括特征偏移和参数偏移。特征移位被识别为由于随机路径采样而引起的隐藏层的动态输入分布。输入分布动态影响损失下降,并最终影响架构排名。参数偏移被识别为在不同训练步骤中针对不同路径中的共享层铺设的矛盾参数更新。快速变化的参数不能保持体系结构的排名。我们同时使用一个非平凡的超网Pi模型,称为Pi-NAS来解决这两个转变。具体来说,我们采用了一个包含交叉路径学习的超网Pi模型,以减少不同路径之间的特征一致性变化。同时,我们采用了一种新的非平凡的平均教师包含负样本,以克服参数移动和模型碰撞。此外,我们的Pi-NAS以无监督的方式运行,可以搜索更多可转移的架构。在ImageNet和广泛的下游任务(例如,COCO 2017、ADE 20 K和Cityscapes)证明了我们的Pi-NAS与监督式NAS相比的有效性和通用性。参见代码:https://github.com/Ernie1/Pi-NAS。

英文摘要:

Recently proposed neural architecture search (NAS) methods co-train billions of architectures in a supernet and estimate their potential accuracy using the network weights detached from the supernet. However, the ranking correlation between the architectures' predicted accuracy and their actual capability is incorrect, which causes the existing NAS methods' dilemma. We attribute this ranking correlation problem to the supernet training consistency shift, including feature shift and parameter shift. Feature shift is identified as dynamic input distributions of a hidden layer due to random path sampling. The input distribution dynamic affects the loss descent and finally affects architecture ranking. Parameter shift is identified as contradictory parameter updates for a shared layer lay in different paths in different training steps. The rapidly-changing parameter could not preserve architecture ranking. We address these two shifts simultaneously using a nontrivial supernet-Pi model, called Pi-NAS. Specifically, we employ a supernet-Pi model that contains cross-path learning to reduce the feature consistency shift between different paths. Meanwhile, we adopt a novel nontrivial mean teacher containing negative samples to overcome parameter shift and model collision. Furthermore, our Pi-NAS runs in an unsupervised manner, which can search for more transferable architectures. Extensive experiments on ImageNet and a wide range of downstream tasks (e.g., COCO 2017, ADE20K, and Cityscapes) demonstrate the effectiveness and universality of our Pi-NAS compared to supervised NAS. See Codes: https://github.com/Ernie1/Pi-NAS.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑