Perceptually Grounded Language Learning: Part 2—DETE: A Neural/Procedural Model

In Part 1 of this two-part series, we introduced Katamic memory—a neural network architecture capable of robust sequence learning and recognition. In Part 2, we introduce the Blobs World taskjdomain for language learning and describe the DETE language learning system, which is composed of over 50 Katamic memory modules. DETE currently learns small subsets of English and Spanish via association with perceptual! motor inputs. In addition to Kaiamic memory, DETE employs several other novel features: (1) use of feature planes, to encode visual shapes, spatial relationships and the motions of objects, (2) phase-locking of neural firing, in order to represent focus of atention and to bind objects across multiple feature planes, and (3) a method for encoding temporal relationships, so that DETE can learn utterances involving the immediate past and future. We compare DETE to related models and discuss the implications of this approach for language-learning research.     

Attentional task aptitude and performance anxiety.

Anxious responding (trait, state, and test anxiety) has a negative impact on overt performance. Researchers have used a unidimensional method of assessing anxiety and performance, although a more informative approach would involve a comprehensive assessment battery and multiple performance tasks. Incorporating this strategy, the authors measured the impact of anxiety on 4 attentional processing tasks. Results revealed that "types" of anxiety symptoms were differentially related to attentional task performance; test anxiety accounted for the most variance in predicting performance on the Wechsler Adult Intelligence Scale-Third Edition (D. Wechsler, 1997) letter-number sequencing and digit-span subtests, trait anxiety and fear of negative evaluation were more significant in predicting Stroop performance, and math anxiety accounted for the largest variance toward understanding Paced Auditory Serial Addition Task (C. W. Lejuez, C. W. Kahler, & R. A. Brown, 2003) scores. Theoretical and clinical implications are discussed. (PsycINFO Database Record (c) 2010 APA, all rights reserved)     

基于损失自注意力机制的立体匹配算法研究

针对现有的立体匹配算法在阴影、物体边缘和光照反射等区域匹配困难且存在大量错误结果的问题,设计了一种可拆卸的损失自注意力网络（loss self-attention net,LSAnet）查找图像中的匹配困难区域。LSAnet的网络各层相互稠密连接,应用了空洞卷积来扩大感受野,并以立体匹配算法生成的损失分布为标签,能够动态地进行有监督训练,最终生成匹配困难区域掩膜辅助立体匹配网络进行更好的优化;同时,改进了立体匹配网络中经典的特征匹配代价卷结构,降低了后续3D卷积的计算负荷,提高了匹配效率。实验结果表明,该算法相比于基准算法精度更高,并且可以提高算法对于匹配困难区域的鲁棒性。     

基于多通路融合网络的高速公路雾天能见度等级识别

雾天是影响高速公路交通安全的重要因素。研究从监控图像进行高速公路雾天能见度的自动识别方法可以为交通管理部门的智能管理和决策提供技术支持。根据大气散射模型分析出与雾浓度相关的多个物理因素,提出了综合这些物理因素的多通路融合识别网络。该网络使用三个通路联合学习深度视觉特征、传输矩阵特征和场景深度特征,并设计注意力融合模块来自适应地融合这三类特征以进行能见度等级识别。同时构建了一个合成数据集和一个真实的高速公路场景数据集,用于网络参数学习和性能评估。实景数据集中的图像是从中国多条高速公路的监控视频中收集的。在这两个数据集上的实验表明,所提方法可以适应不同的监控拍摄场景,能够比现有方法更准确地识别能见度等级,有效提升了识别精度。     

局部特征表征的6D位姿估计算法

为解决有纹理模型在遮挡条件下6D位姿估计精确度不高的问题,提出了一种局部特征表征的端到端6D位姿估计算法。首先为了得到准确的定位信息,提出了一个空间—坐标注意力机制（spatial and coordinate attention）,通过在YOLOv5网络中加入空间—坐标注意力机制和加权双向特征金字塔网络（bidirectional feature pyramid network）,YOLOv5-CBE算法的精确度（precision）、召回率（recall）、平均精度均值（mAP@0.5）分别提升了3.6%、2.8%、2.5%,局部特征中心点坐标误差最高提升了25%;然后用 YOLOv5-CBE算法检测局部特征关键点,结合3D Harris关键点通过奇异值分解法（singular value decomposition）计算模型的6D位姿,最高遮挡70%的情况下仍然可以保证二维重投影精度（2D reprojection accuracy）和ADD度量精度（ADD accuracy）在95%以上,具有较强的鲁棒性。     

基于标签推理和注意力融合的多标签文本分类方法

目前许多多标签文本分类方法主要关注文档表示,而丢失了大量标签相关的语义信息,导致分类效果不理想。针对以上问题,提出一种基于标签推理和注意力融合的分类方法,挖掘文档中与标签相关的特征以及相似标签之间的相关性,学习标签信息进行标签推理,同时采用注意力机制自学习地融合文档表示和标签表示,最终完成多标签分类任务。在AAPD和RCV1-V2数据集上进行实例验证,该方法的F1值分别达到了0.732和0.887,与其他最新方法相比其准确度均有提升,实验结果证明了标签推理和注意力融合策略的有效性。     

基于组反馈融合机制的视频超分辨率模型

视频超分辨率（video super-resolution,VSR）的目的是利用多个相邻帧的信息来生成参考帧的高分辨率版本。现有的许多VSR工作都集中在如何有效地对齐相邻帧以更好地融合相邻帧信息,而很少在相邻帧信息融合这一重要步骤上进行研究。针对该问题,提出了基于组反馈融合机制的视频超分辩模型（GFFMVSR）。具体来说,在相邻帧对齐后,将对齐视频序列输入第一重时间注意力模块;然后,将序列分成几个小组,各小组依次通过组内融合模块实现初步融合。不同小组的融合结果经过第二重时间注意力模块;然后,各小组逐组输入反馈融合模块,利用反馈机制反馈融合不同组别的信息,最后将融合结果输出重建。经验证,该模型具有较强的信息融合能力,在客观评价指标和主观视觉效果上都优于现有的模型。     

Quadratic polynomial guided fuzzy C-means and dual attention mechanism for medical image segmentation

Medical image segmentation is the most complex and important task in the field of medical image processing and analysis, as it is linked to disease diagnosis accuracy. However, due to the medical image's high complexity and noise, segmentation performance is limited. We propose a novel quadratic polynomial guided fuzzy C-means and dual attention mechanism composite network model architecture to address the aforementioned issues (QPFC-DA). It has mechanisms for channel and spatial edge attention, which guide the content and edge segmentation branches, respectively. The bi-directional long short-term memory network was added after the two content segmentation branches to better integrate multi-scale features and prevent the loss of important features. Furthermore, the fuzzy C-means algorithm guided by the quadratic polynomial can better distinguish the image's weak edge regions and has a degree of noise resistance, resulting in a membership matrix with less ambiguity and a more reliable segmentation result. We also conducted comparison and ablation experiments on three medical data sets. The experimental results show that this method is superior to several other well-known methods.     

CoGCN: Combining co‐attention with graph convolutional network for entity linking with knowledge graphs

Entity linking is a fundamental task in natural language processing. The task of entity linking with knowledge graphs aims at linking mentions in text to their correct entities in a knowledge graph like DBpedia or YAGO2. Most of existing methods rely on hand‐designed features to model the contexts of mentions and entities, which are sparse and hard to calibrate. In this paper, we present a neural model that first combines co‐attention mechanism with graph convolutional network for entity linking with knowledge graphs, which extracts features of mentions and entities from their contexts automatically. Specifically, given the context of a mention and one of its candidate entities' context, we introduce the co‐attention mechanism to learn the relatedness between the mention context and the candidate entity context, and build the mention representation in consideration of such relatedness. Moreover, we propose a context‐aware graph convolutional network for entity representation, which takes both the graph structure of the candidate entity and its relatedness with the mention context into consideration. Experimental results show that our model consistently outperforms the baseline methods on five widely used datasets.     

一种时空注意力网络的交通预测模型

交通预测在智能交通中有着重要的意义和应用.由于交通数据的复杂性和高度的非线性,精确的交通预测的核心挑战在于如何对复杂的空间相关性和时间动态建立模型.在现实生活中,我们发现:1)区域间的空间依赖是动态的;2)时间依赖有日和周的模式,但由于有动态时间变化,它不具有严格周期性.为了解决这两个问题,我们提出了一个新的时空注意力网络(STAN),该模型的主要思想是区域间的动态相似性用一个门控机制学习,长期周期性时间转移现象由一个周期性注意力转移机制来学习,并考虑交通道路、天气状况等外部因素.通过与不同的方法在两个数据集上进行评估,实验结果表明,我们提出的模型有更好的准确性.