Decision support in machine vision system for monitoring of TFT-LCD glass substrates manufacturing

This study addresses classification methodology for the automatic inspection of a range of defects on the surface of glass substrates in thin film transistor liquid crystal display glass substrate manufacturing. The proposed methodology consisted of four stages: (1) feature extraction by calculating the wavelet co-occurrence signature from the substrate images, (2) handling of imbalanced dataset using the Synthetic Minority Over-sampling TEchnique (SMOTE), (3) reduction of the feature's dimension by principal component analysis, and (4) finally choosing the best classifier between three different methods: Classification And Regression Tree (CART), Multi-Layer Perceptron (MLP) and Support Vector Machine (SVM). In training the SVM and MLP classifiers, the simulated annealing algorithm was used to obtain the optimal tuning parameters for the classifiers. From the industrial case study, the proposed feature extraction algorithm could remove the defect-irrelevant image features and SMOTE increased the accuracy of all three methods. Furthermore, the optimized SVM and MLP models were more accurate than the CART model whereas a higher accuracy of 89.5% was observed for the proposed SVM model.     

Multimodal Pre-training Method for Vision-language Understanding and Generation

Most existing vision-language pre-training methods focus on understanding tasks and use BERT-like loss functions (masked language modeling and image-text matching) during pre-training. Despite their good performance in the understanding of downstream tasks, such as visual question answering, image-text retrieval, and visual entailment, these methods cannot generate information. To tackle this problem, this study proposes Unified multimodal pre-training for Vision-Language understanding and generation (UniVL). The proposed UniVL is capable of handling both understanding tasks and generation tasks. It expands existing pre-training paradigms and uses random masks and causal masks simultaneously, where causal masks are triangular masks that mask future tokens, and such pre-trained models can have autoregressive generation abilities. Moreover, several vision-language understanding tasks are turned into text generation tasks according to specifications, and the prompt-based method is employed for fine-tuning of different downstream tasks. The experiments show that there is a trade-off between understanding tasks and generation tasks when the same model is used, and a feasible way to improve both tasks is to use more data. The proposed UniVL framework attains comparable performance to recent vision-language pre-training methods in both understanding tasks and generation tasks. Moreover, the prompt-based generation method is more effective and even outperforms discriminative methods in few-shot scenarios.     

基于深度学习的语言模型研究进展

语言模型旨在对语言的内隐知识进行表示,作为自然语言处理的基本问题,一直广受关注．基于深度学习的语言模型是目前自然语言处理领域的研究热点,通过预训练-微调技术展现了内在强大的表示能力,并能够大幅提升下游任务性能．本文围绕语言模型基本原理和不同应用方向,以神经概率语言模型与预训练语言模型作为深度学习与自然语言处理结合的切入点,从语言模型的基本概念和理论出发,介绍了神经概率与预训练模型的应用情况和当前面临的挑战,对现有神经概率、预训练语言模型及方法进行对比和分析．我们又从新型训练任务和改进网络结构两方面对预训练语言模型训练方法进行详细阐述,并对目前预训练模型在规模压缩、知识融合、多模态和跨语言等研究方向进行概述和评价．最后总结语言模型在当前自然语言处理应用中的瓶颈,对未来可能的研究重点做出展望．     

基于多任务预训练的AMR文本生成研究

抽象语义表示（abstract meaning representation,简称AMR）文本生成的任务是给定AMR图,生成与其语义一致的文本.相关工作表明,人工标注语料的规模大小直接影响了AMR文本生成的性能.为了降低对人工标注语料的依赖,提出了基于多任务预训练的AMR文本生成方法.特别地,基于大规模自动标注AMR语料,提出与AMR文本生成任务相关的3个预训练任务,分别是AMR降噪自编码、句子降噪自编码以及AMR文本生成任务本身.此外,基于预训练模型,在朴素微调方法的基础上,进一步提出了基于多任务训练的微调方法,使得最终模型不仅适用于AMR文本生成,同时还适用于预训练任务.基于两个AMR标准数据集的实验结果表明：使用0.39M自动标注数据,提出的预训练方法能够大幅度提高AMR文本生成的性能,在AMR2.0和AMR3.0上分别提高了12.27和7.57个BLEU值,性能分别达到40.30和38.97.其中,在AMR2.0上的性能为目前报告的最优值,在AMR3.0上的性能为目前为止首次报告的性能.     

文本情感分析方法研究综述

文本情感分析是自然语言处理领域的一个重要分支,广泛应用于舆情分析和内容推荐等方面,是近年来的研究热点。根据使用的不同方法,将其划分为基于情感词典的情感分析方法、基于传统机器学习的情感分析方法、基于深度学习的情感分析方法。通过对这三种方法进行对比,分析其研究成果,并对不同方法的优缺点进行归纳总结,介绍相关数据集和评价指标及应用场景,对情感分析子任务进行简单概括,发现将来的情感分析问题的研究趋势及应用领域,并为研究者在相关领域方面提供一定的帮助和指导。     

Supervised Distance Preserving Projections: Applications in the quantitative analysis of diesel fuels and light cycle oils from NIR spectra

In this work, we discuss a recently proposed approach for supervised dimensionality reduction, the Supervised Distance Preserving Projection (SDPP) and, we investigate its applicability to monitoring material's properties from spectroscopic observations. Motivated by continuity preservation, the SDPP is a linear projection method where the proximity relations between points in the low-dimensional subspace mimic the proximity relations between points in the response space. Such a projection facilitates the design of efficient regression models and it may also uncover useful information for visualisation. An experimental evaluation is conducted to show the performance of the SDPP and compare it with a number of state-of-the-art approaches for unsupervised and supervised dimensionality reduction. The regression step after projection is performed using computationally light models with low maintenance cost like Multiple Linear Regression and Locally Linear Regression with k-NN neighbourhoods. For the evaluation, a benchmark and a full-scale calibration problem are discussed. The case studies pertain the estimation of a number of chemico-physical properties in diesel fuels and in light cycle oils, starting from near-infrared spectra. Based on the experimental results, we found that the SDPP leads to parsimonious projections that can be used to design light and yet accurate estimation models.     

基于监督协同近邻保持投影的人脸识别算法

基于流形学习理论的近邻保持嵌入算法(Neighborhood Preserving Embedding,NPE)能够发现数据集中隐含的内蕴结构,但当训练样本不足时,无法准确发现数据的内在流形结构,从而影响算法的识别效果.针对这一问题,对NPE算法进行改进,提出了监督协同近邻保持投影算法(Supervised Collaborative Neighborhood Preserving Projection,SCNPP).该算法在类别信息的指导下构建近邻图,使同类样本间的几何关系得到保持,利用协同表示弥补NPE因样本不足造成的表示误差,以一个有效保持样本近邻关系、准确发现数据内在流形结构的权值矩阵计算投影矩阵,提高分类效果.在FERET、AR和Extended Yale B人脸数据集上的实验验证了该算法的有效性.     

Supervised classification of share price trends

Share price trends can be recognized by using data clustering methods. However, the accuracy of these methods may be rather low. This paper presents a novel supervised classification scheme for the recognition and prediction of share price trends. We first produce a smooth time series using zero-phase filtering and singular spectrum analysis from the original share price data. We train pattern classifiers using the classification results of both original and filtered time series and then use these classifiers to predict the future share price trends. Experiment results obtained from both synthetic data and real share prices show that the proposed method is effective and outperforms the well-known K-means clustering algorithm.     

Transaction aggregation as a strategy for credit card fraud detection

The problem of preprocessing transaction data for supervised fraud classification is considered. It is impractical to present an entire series of transactions to a fraud detection system, partly because of the very high dimensionality of such data but also because of the heterogeneity of the transactions. Hence, a framework for transaction aggregation is considered and its effectiveness is evaluated against transaction-level detection, using a variety of classification methods and a realistic cost-based performance measure. These methods are applied in two case studies using real data. Transaction aggregation is found to be advantageous in many but not all circumstances. Also, the length of the aggregation period has a large impact upon performance. Aggregation seems particularly effective when a random forest is used for classification. Moreover, random forests were found to perform better than other classification methods, including SVMs, logistic regression and KNN. Aggregation also has the advantage of not requiring precisely labeled data and may be more robust to the effects of population drift.