增量任务学习（ITL）是一个持续学习的类别，试图培训单个网络以进行多个任务（一个接一个），其中每个任务的培训数据仅在培训该任务期间可用。当神经网络接受较新的任务培训时，往往会忘记旧任务。该特性通常被称为灾难性遗忘。为了解决此问题，ITL方法使用情节内存，参数正则化，掩盖和修剪或可扩展的网络结构。在本文中，我们提出了一个基于低级别分解的新的增量任务学习框架。特别是，我们表示每一层的网络权重作为几个等级1矩阵的线性组合。为了更新新任务的网络，我们学习一个排名1（或低级别）矩阵，并将其添加到每一层的权重。我们还引入了一个其他选择器向量，该向量将不同的权重分配给对先前任务的低级矩阵。我们表明，就准确性和遗忘而言，我们的方法的表现比当前的最新方法更好。与基于情节的内存和基于面具的方法相比，我们的方法还提供了更好的内存效率。我们的代码将在https://github.com/csiplab/task-increment-rank-update.git上找到。

现有的唱歌语音合成模型（SVS）通常在唱歌数据上进行训练，并取决于容易出错的时间对齐和持续时间功能或明确的音乐得分信息。在本文中，我们提出了Karaoker，Karaoker是一种基于多言式Tacotron的模型，该模型以语音特征为条件，该功能专门针对口语数据进行训练，而无需时间对齐。卡拉克（Karaoker）在从看不见的歌手/扬声器的源波形中提取的多维模板之后，综合了歌声和传输风格。该模型在连续数据上以单个深卷积编码为共同条件，包括音高，强度，和谐，实扣，cepstral峰值突出和八度。我们通过功能重建，分类和说话者身份识别任务扩展了文本到语音训练目标，这些任务将模型指导到准确的结果。除多任务外，我们还采用了Wasserstein GAN训练方案以及声学模型的输出的新损失，以进一步完善模型的质量。

本文介绍了对F0的音素级韵律控制的方法和多销箱文本到语音设置的持续时间，基于韵律聚类。使用自回归关注的模型，并将多个箱子架构模块并联，与韵律编码器并联。提出了对基本单扬声器方法的几种改进，从而增加了韵律控制范围和覆盖范围。更具体地说，我们采用数据增强，F0​​标准化，持续时间的平衡集群，以及扬声器无关的韵律聚类。这些修改使培训集中包含的所有发言者能够进行细粒度的音素级韵律控制，同时保持扬声器标识。该模型也可以微调到具有限制数据量的看不见的扬声器，并显示其维持其韵律控制能力，验证说话者无关的韵律聚类是有效的。实验结果验证了该模型维持了高输出语音质量，并且该方法允许在每个扬声器范围内有效的韵律控制，尽管多种式箱子设置介绍的变化。

在本文中，介绍了文本到读取/唱歌系统，可以适应任何扬声器的声音。它利用基于TacoTron的多级箱子声学模型在只读语音数据训练，并且在音素级别提供韵律控制。还研究了基于传统DSP算法的数据集增强和额外的韵律操纵。神经TTS模型对看不见的扬声器的有限录音进行了微调，允许与目标的扬声器语音进行敲击/歌唱合成。描述了系统的详细管道，其包括从Capella歌曲的目标音调和持续时间值提取，并将其转换为在合成之前的目标扬声器的有效音符范围内。还研究了通过WSOLA输出的输出的韵律操纵的另外的阶段，以便更好地匹配目标持续时间值。合成的话语可以与乐器伴奏轨道混合以产生完整的歌曲。通过主观聆听测试评估所提出的系统，以及与可用的备用系统相比，该系统还旨在从只读训练数据产生合成歌唱语音。结果表明，该拟议的方法可以产生高质量的敲击/歌声，具有增加的自然。

本文介绍了一个端到端的文本到语音系统，CPU延迟低，适用于实时应用。该系统由基于自回归关注的序列到序列声学模型和用于波形生成的LPCNet声码器组成。提出了一种采用塔克罗伦1和2型号的模块的声学模型架构，而通过使用最近提出的基于位置的注意机制来确保稳定性，适用于任意句子长度。在推断期间，解码器是展开的，并且以流式方式执行声学特征生成，允许与句子长度无关的几乎恒定的延迟。实验结果表明，声学模型可以产生比计算机CPU上的实时大约31倍的功能序列，移动CPU上的6.5倍，使其能够满足两个设备上实时应用所需的条件。全端到端系统可以通过听证测试来验证几乎是自然的质量语音。

A methodology is proposed, which addresses the caveat that line-of-sight emission spectroscopy presents in that it cannot provide spatially resolved temperature measurements in nonhomogeneous temperature fields. The aim of this research is to explore the use of data-driven models in measuring temperature distributions in a spatially resolved manner using emission spectroscopy data. Two categories of data-driven methods are analyzed: (i) Feature engineering and classical machine learning algorithms, and (ii) end-to-end convolutional neural networks (CNN). In total, combinations of fifteen feature groups and fifteen classical machine learning models, and eleven CNN models are considered and their performances explored. The results indicate that the combination of feature engineering and machine learning provides better performance than the direct use of CNN. Notably, feature engineering which is comprised of physics-guided transformation, signal representation-based feature extraction and Principal Component Analysis is found to be the most effective. Moreover, it is shown that when using the extracted features, the ensemble-based, light blender learning model offers the best performance with RMSE, RE, RRMSE and R values of 64.3, 0.017, 0.025 and 0.994, respectively. The proposed method, based on feature engineering and the light blender model, is capable of measuring nonuniform temperature distributions from low-resolution spectra, even when the species concentration distribution in the gas mixtures is unknown.

The two popular datasets ScanRefer [16] and ReferIt3D [3] connect natural language to real-world 3D data. In this paper, we curate a large-scale and complementary dataset extending both the aforementioned ones by associating all objects mentioned in a referential sentence to their underlying instances inside a 3D scene. Specifically, our Scan Entities in 3D (ScanEnts3D) dataset provides explicit correspondences between 369k objects across 84k natural referential sentences, covering 705 real-world scenes. Crucially, we show that by incorporating intuitive losses that enable learning from this novel dataset, we can significantly improve the performance of several recently introduced neural listening architectures, including improving the SoTA in both the Nr3D and ScanRefer benchmarks by 4.3% and 5.0%, respectively. Moreover, we experiment with competitive baselines and recent methods for the task of language generation and show that, as with neural listeners, 3D neural speakers can also noticeably benefit by training with ScanEnts3D, including improving the SoTA by 13.2 CIDEr points on the Nr3D benchmark. Overall, our carefully conducted experimental studies strongly support the conclusion that, by learning on ScanEnts3D, commonly used visio-linguistic 3D architectures can become more efficient and interpretable in their generalization without needing to provide these newly collected annotations at test time. The project's webpage is https://scanents3d.github.io/ .

Natural language interaction is a promising direction for democratizing 3D shape design. However, existing methods for text-driven 3D shape editing face challenges in producing decoupled, local edits to 3D shapes. We address this problem by learning disentangled latent representations that ground language in 3D geometry. To this end, we propose a complementary tool set including a novel network architecture, a disentanglement loss, and a new editing procedure. Additionally, to measure edit locality, we define a new metric that we call part-wise edit precision. We show that our method outperforms existing SOTA methods by 20% in terms of edit locality, and up to 6.6% in terms of language reference resolution accuracy. Our work suggests that by solely disentangling language representations, downstream 3D shape editing can become more local to relevant parts, even if the model was never given explicit part-based supervision.

The current success of machine learning on image-based combustion monitoring is based on massive data, which is costly even impossible for industrial applications. To address this conflict, we introduce few-shot learning in order to achieve combustion monitoring and classification for the first time. Two algorithms, Siamese Network coupled with k Nearest Neighbors (SN-kNN) and Prototypical Network (PN), were tested. Rather than utilizing solely visible images as discussed in previous studies, we also used Infrared (IR) images. We analyzed the training process, test performance and inference speed of two algorithms on both image formats, and also used t-SNE to visualize learned features. The results demonstrated that both SN-kNN and PN were capable to distinguish flame states from learning with merely 20 images per flame state. The worst performance, which was realized by PN on IR images, still possessed precision, accuracy, recall, and F1-score above 0.95. We showed that visible images demonstrated more substantial differences between classes and presented more consistent patterns inside the class, which made the training speed and model performance better compared to IR images. In contrast, the relatively low quality of IR images made it difficult for PN to extract distinguishable prototypes, which caused relatively weak performance. With the entrire training set supporting classification, SN-kNN performed well with IR images. On the other hand, benefitting from the architecture design, PN has a much faster speed in training and inference than SN-kNN. The presented work analyzed the characteristics of both algorithms and image formats for the first time, thus providing guidance for their future utilization in combustion monitoring tasks.

地球天文台是一个不断增长的研究领域，可以在短时间预测（即现在的情况下）利用AI的力量。在这项工作中，我们使用视频变压器网络应对天气预报的挑战。视觉变压器体系结构已在各种应用中进行了探索，主要限制是注意力的计算复杂性和饥饿的培训。为了解决这些问题，我们建议使用视频Swin-Transformer，再加上专用的增强计划。此外，我们在编码器侧采用逐渐的空间减少，并在解码器上进行了交叉注意。在Weather4cast2021天气预报挑战数据中测试了建议的方法，该数据需要从每小时的天气产品序列预测未来的8小时（每小时4个小时）。将数据集归一化为0-1，以促进使用不同数据集的评估指标。该模型在提供训练数据时会导致MSE得分为0.4750，在不使用培训数据的情况下转移学习过程中为0.4420。