点击率(CTR)预测一直是计算广告中最核心的问题之一。最近,生成广告ID的低维表示的嵌入技术极大地提高了CTR预测的准确性。然而,这种学习技术对于具有很少记录数据的新广告来说是数据要求和工作量很大,这被称为冷启动问题。在本文中,我们的目标是在新的广告添加到候选池时,在冷启动阶段和预热阶段期间改进CTR预测。我们提出了Meta-Embedding,这是一种基于元学习的方法,可以学习为新的广告ID生成合适的初始嵌入。所提出的方法通过利用先前学习的adsthrough基于梯度的元学习来训练用于新广告ID的联合生成器。换句话说,我们的方法可以学习如何更好地嵌入。当新广告到来时,受过训练的生成器通过提供其内容和属性来初始化其ID的嵌入。接下来,与现有的初始化方法相比,生成的嵌入可以在一些标记的示例可用时在加热阶段加速模型拟合。三个真实世界数据集的实验结果表明,Meta-Embedding可以显着提高六种现有CTR预测模型的冷启动和预热性能,从轻量级模型(如Factorization Machines)到复杂的深度模型(如PNN和DeepFM)。以上所有内容也适用于转换率(CVR)预测。
translated by 谷歌翻译
在本研究中,我们专注于从Web语料库中提取知识渊博的片段和注释可知文档,其中包括来自社会媒体和We-media的文档。非正式地,知识渊博的片段是指文本描述概念,实体的属性或实体之间的关系,而知识文档是具有足够知识的片段的文档。这些可知的片段和文档可以在多种应用中有所帮助,例如知识库构建和面向知识的服务。以前的研究使用基于模式的方法提取了知识渊博的片段。在这里,我们提出了基于语义的方法来完成这项任务。具体而言,开发基于CNN的模型以同时提取知识渊博的片段和注释可知文档。此外,CNN的“低级共享,高级别拆分”结构旨在处理来自不同内容域的文档。与构建多个特定领域的CNN相比,该联合模型不仅可以大大节省训练时间,而且可以明显提高预测精度。在Wechat公共平台的真实数据集中演示了所提出的方法的优越性。
translated by 谷歌翻译
数据分类存在于不同的实际问题中,例如识别图像中的图案,区分生产线中的缺陷部分,对良性和恶性肿瘤进行分类等。这些问题中的许多都具有难以识别的数据模式,这需要更先进的分辨技术。最近,已经应用了几种针对不同人工神经网络架构的工作来解决分类问题。当分类问题必须通过图像获得时,目前,标准方法是使用卷积神经网络。因此,在本报告中,卷积神经网络被用来对鱼类进行分类。 Classifica \ c {c} \〜ao de dados est \'a presente em diversos problemas reais,tais como:reconhecer padr \〜oes em imagens,diferenciar pe \ c {c} as defeituosasem uma linha de produ \ c {c} \〜ao,classificar tumores benignos e malignos,dentrediversas outras。 Muitos认为问题可能是错误的问题,他们可能会发现问题,但是他们会在这里找到问题,他们会在这里找到自己的想法。 Recentemente,diversos trabalhosabordando diferentes arquiteturas de redes neurais artificiais v \ ^ em sendoaplicados para solucionar problemas de classifica \ c {c} \ ~ao。 Quando aclassifica \ c {c} \〜ao do problema deve ser obtida por meio de imagens,atualmentea metodologia padr \〜ao \'e udes de redes neurais convolucionais。 Sendo assim,neste trabalho s \〜ao utilizadas redes neurais convolucionais paraclassifica \ c {c} \ ~ao de esp \'ecies de peixes。
translated by 谷歌翻译
股骨近端骨折是西方世界的一个重要实体,特别是随着老年人口的增长。这种骨折导致高发病率和死亡率,反映了对我们社会的重大健康和经济影响。对于不同的骨折类型,建议采用不同的治疗策略,手术治疗仍是大多数病例的金标准。在外科手术后治疗和预后的成功取决于标准类型之间的骨折的准确分类,例如由AO系统定义的那些。然而,基于X射线图像的分类断裂类型是困难的,这通过我们内部研究的低内部和专家间协议率以及之前的文献证实。所提出的工作提出了一种基于当前深度学习技术的全自动计算机辅助诊断(CAD)工具,能够根据AO分类识别,定位并最终对X射线图像上的股骨近端骨折进行分类。我们的实验评估结果表明,所提出的CAD工具所达到的性能可与普通专家相比,对X射线图像进行分类,即“A”,“B”和“正常”(精度为89%) ),虽然在将骨折分类与“正常”病例(精确度为94%)时表现优异。此外,广泛讨论了将所提出的CAD工具集成到日常临床常规中,以改善人类和人工智能机器之间的界面,以支持医疗决策。
translated by 谷歌翻译
近年来,随着气候变化效应和世界人口增长,对生产力,精度和效率的高要求,机器人在农业中的使用一直在增加。与常规农业不同,甘蔗农场通常是植被茂密,面积巨大的地区,并且受到极端天气条件的影响,如剧烈的热量,潮湿和雨水。 TIBA - 用于智能生物能源农业的Tankette--是研发项目的第一个成果,该项目致力于开发一个自动移动机器人系统,用于执行许多农业任务,以实现许多农业领域。提出的概念包括半自动,低成本,防尘和防水罐车型车辆,能够渗透人工林隧道中的茂密植被并携带多个传感系统,以便对难以进入的区域进行绘图并收集样本。本文介绍了机器人机械设计,嵌入式电子设备和软件体系结构以及第一个原型的构建。在现场测试中获得的初步结果验证了所提出的概念设计,并为机器人自主导航带来了若干挑战和潜在应用,以及构建具有附加功能的新原型。
translated by 谷歌翻译
使用神经网络直接预测用于自动规划的三维分配分布正变得流行。然而,现有方法仅使用患者解剖结构作为输入并且对训练数据库中的所有患者采取一致的波束配置。这项工作的目的是开发一个更通用的模型,除了患者解剖学,还考虑可变光束配置,以实现更全面的自动规划与可能更容易的临床实施,而无需针对不同的光束设置训练特定模型。
translated by 谷歌翻译
变化检测涉及分割顺序数据,使得相同分段中的观察共享一些期望的属性。多变量检测仍然是一个具有挑战性的问题,因为各种方式的转换点可以在各个通道之间进行相关,并且各个通道上的潜在信噪比也是如此。在本文中,我们感兴趣的是在无监督设置中定位加性异常值(AO)和水平移位(LS)。我们提出ABACUS,自动BAyesian Changepoints Under Sparsity,aBayesian源分离技术,用于恢复潜在信号,同时还可以检测模型参数的变化。多级稀疏性实现了尺寸减小和信号变化建模。我们展示了ABACUS在模拟研究中具有竞争力或卓越性能,可以应对最先进的变化检测方法和已建立的潜在变量模型。我们还介绍了ABACUS在两个实际应用,建模基因组配置文件和分析家庭用电量。
translated by 谷歌翻译
This study investigates the hypothesis that speakers make active use of the visual modality in production to improve their speech intelligibility in noisy conditions. Six native speakers of Canadian French produced speech in quiet conditions and in 85 dB of babble noise, in three situations: interacting face-to-face with the experimenter (AV), using the auditory modality only (AO), or reading aloud (NI, no interaction). The audio signal was recorded with the three-dimensional movements of their lips and tongue, using electromagnetic articulography. All the speakers reacted similarly to the presence vs absence of communicative interaction, showing significant speech modifications with noise exposure in both interactive and non-interactive conditions, not only for parameters directly related to voice intensity or for lip movements (very visible) but also for tongue movements (less visible); greater adaptation was observed in interactive conditions, though. However, speakers reacted differently to the availability or unavailability of visual information: only four speakers enhanced their visible articulatory movements more in the AV condition. These results support the idea that the Lombard effect is at least partly a listener-oriented adaptation. However, to clarify their speech in noisy conditions, only some speakers appear to make active use of the visual modality .
translated by 谷歌翻译
Figure 1: Left: Direct and hemispherical ambient illumination in San Miguel (6.5M triangles, 968 draw calls). Right: Direct lighting, approximate radiosity, mirror reflections, and AO computed from a two-layer Deep G-buffer in 5 ms at 1080p on NVIDIA GeForce 980. The G-buffer was generated in a single 5.8ms geometry pass. See our evaluation section for faster results on more game-like scenes. Abstract We introduce a new hardware-accelerated method for constructing Deep G-buffers that is 2x-8x faster than the previous depth peeling method and produces more stable results. We then build several high-performance shading algorithms atop our representation , including dynamic diffuse interreflection, ambient occlusion (AO), and mirror reflection effects. Our construction method s order-independent, guarantees a minimum separation between layers, operates in a (small) bounded memory footprint, and does not require per-pixel sorting. Moreover, addressing the increasingly expensive cost of pre-rasterization, our approach requires only a single pass over the scene geometry. Our global illumination algorithms approach the speed of the fastest screen-space AO-only techniques while significantly exceeding their quality: we capture small-scale details and complex radiometric effects more robustly than screen-space techniques, and we implicitly handle dynamic illumination conditions. We include the pseudocode for our Deep G-buffer construction in the paper and the full source code of our technique in our supplemental document.
translated by 谷歌翻译
To predict future changes in blocking and the resulting weather extremes, some studies have proposed the negative phase of Arctic Oscillation (−AO) as an analogue for Arctic amplification because of similarities between their mean states: reduced midlatitude-to-pole temperature gradients and weakened, equatorward shifted jet streams. Using well-controlled modeling experiments, we show that blocking variations associated with mean state anomalies are opposite depending on whether these anomalies are driven by the internal dynamics as in AO or forced externally as in Arctic amplification. While blocking increases and its latitudinal-distribution shifts poleward in −AO, we find opposite responses when a mean state identical to the −AO mean state is externally forced. Findings suggest that the observed blocking-AO relationship is a correlation which does not imply that the −AO mean state causes increased blocking and should not be employed as a prototype for Arctic amplification. Furthermore, results urge for a careful consideration of causality before using internal variability to predict low-frequency response to external forcings.
translated by 谷歌翻译