人工智能培训

搜索

人工智能论文:稀疏局部密集液体氩时间投影室数据的可扩展深卷积神经网络(Scalable Deep Convolutional Neural Networks f

[复制链接]
trueno 发表于 2019-3-15 12:02:40 | 显示全部楼层 |阅读模式
trueno 2019-3-15 12:02:40 121 0 显示全部楼层
人工智能论文:稀疏局部密集液体氩时间投影室数据的可扩展深卷积神经网络(Scalable Deep Convolutional Neural Networks for Sparse, Locally Dense  Liquid Argon Time Projection Chamber Data)深度卷积神经网络(CNN)显示出在许多领域中分析科学数据的巨大前景,包括粒子成像探测器,如液体氩时间投影室(LArTPC)。然而,LArTPCdata的高稀疏性挑战了传统的CNN,这些CNN是专为密集数据设计的。 CNN对LArTPC数据的简单应用导致低效计算和大型LArTPC检测器(例如ShortBaseline中微子程序和深地下中微子实验)的可扩展性差。最近,已经提出了子流量稀疏卷积网络(SSCN)来应对这一挑战。我们在模拟的LArTPC样本上报告了他们在3D语义分割任务上的表现。与标准CNN相比,我们观察到推理的计算存储器和壁时间成本分别降低了364和33倍,而没有损失精度。 2D样本的相同因子分别为93和3.1。使用SSCN,我们提出了使用公共3D LArTPC样本重建Michel电子的第一种基于机器学习的方法。我们发现Michel电子识别效率为93.9%,真阳性率为98.8%。重建的Michelelectron簇的平均像素聚类效率为96.1%,纯度为97.3%。结果令人信服地展示了使用深度神经网络进行大规模LArTPC探测器的可扩展数据重建技术的强大前景。
Deep convolutional neural networks (CNNs) show strong promise for analyzingscientific data in many domains including particle imaging detectors such as aliquid argon time projection chamber (LArTPC).Yet the high sparsity of LArTPCdata challenges traditional CNNs which were designed for dense data such asphotographs.A naive application of CNNs on LArTPC data results in inefficientcomputations and a poor scalability to large LArTPC detectors such as the ShortBaseline Neutrino Program and Deep Underground Neutrino Experiment.RecentlySubmanifold Sparse Convolutional Networks (SSCNs) have been proposed to addressthis challenge.We report their performance on a 3D semantic segmentation taskon simulated LArTPC samples.In comparison with standard CNNs, we observe thatthe computation memory and wall-time cost for inference are reduced by factorof 364 and 33 respectively without loss of accuracy.The same factors for 2Dsamples are found to be 93 and 3.1 respectively.Using SSCN, we present thefirst machine learning-based approach to the reconstruction of Michel electronsusing public 3D LArTPC samples.We find a Michel electron identificationefficiency of 93.9\% with 98.8\% of true positive rate.Reconstructed Michelelectron clusters yield 96.1\% in average pixel clustering efficiency and97.3\% in purity.The results are compelling to show strong promise of scalabledata reconstruction technique using deep neural networks for large scale LArTPCdetectors.人工智能论文:稀疏局部密集液体氩时间投影室数据的可扩展深卷积神经网络(Scalable Deep Convolutional Neural Networks for Sparse, Locally Dense  Liquid Argon Time Projection Chamber Data) nuZpvYQ3u9IcvqYL.jpg
URL地址:https://arxiv.org/abs/1903.05663     ----pdf下载地址:https://arxiv.org/pdf/1903.05663    ----人工智能论文:稀疏局部密集液体氩时间投影室数据的可扩展深卷积神经网络(Scalable Deep Convolutional Neural Networks for Sparse, Locally Dense  Liquid Argon Time Projection Chamber Data)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

trueno当前离线
新手上路

查看:121 | 回复:0

快速回复 返回顶部 返回列表