Anomaly Detection within High Volume Data Streams Using Deep Learning Industrial Partner: Cosmonio Ltd (http://cosmonio.com/)

使用深度学习在大量数据流中进行异常检测工业合作伙伴：Cosmonio Ltd (http://cosmonio.com/)

• 批准号: 1966980
• 金额: --
• 依托单位: Durham University
• 依托单位国家: 英国
• 项目类别: Studentship
• 财政年份: 2017
• 资助国家: 英国
• 起止时间: 2017 至 无数据
• 项目状态: 已结题
• 来源: https://gtr.ukri.org/projects?ref=studentship-1966980
• 关键词: Anomaly; Detection; within; Volume; Data

Deep Learning is a fast-growing area in the field of machine learning within which systems can be trained to identify patterns in a way that loosely represent biological neural computation. Recent advances in such Deep Neural Network (DNN) technology have led to a significant performance leap in pattern-recognition tasks such as computer vision (image understanding), text mining (languageunderstanding) and voice recognition (audio understanding). This advancement has been further enhanced by the evolution of low-cost Graphical Processing Units (GPUs) into massively-parallel computing platforms capable of processing the vast dataset requirements for such tasks.This project looks beyond the initial applications of such technology, largely aimed at the classification of images, words and sound to a set of expected semantic labels to the problem of anomaly detection.Asking the novel research question - what is different from normal here within this stream of data? - rather than specifically identifying the {pixel | text | word ...} pattern for {person | vehicle | dog ...} etc in the conventional sense.

深度学习是机器学习领域中一个快速发展的领域，在这个领域中，可以训练系统以松散地表示生物神经计算的方式识别模式。这种深度神经网络（DNN）技术的最新进展导致了模式识别任务的显著性能飞跃，例如计算机视觉（图像理解），文本挖掘（语言理解）和语音识别（音频理解）。这一进步已经进一步加强了低成本的图形处理单元（GPU）的发展成为并行计算平台，能够处理大量的数据集的要求，这样的任务。该项目着眼于超越这种技术的最初应用，主要是针对图像的分类，单词和声音到一组预期的语义标签的异常检测的问题。问新的研究问题-什么是不同的正常在这个数据流在这里？- 而不是具体地标识{像素|文本|字.} {person}的模式|车辆|狗...等传统意义上的。

项目成果

Autoencoders Without Reconstruction for Textural Anomaly Detection

• DOI: 10.1109/ijcnn52387.2021.9533804
• 发表时间: 2021-07
• 期刊: 2021 International Joint Conference on Neural Networks (IJCNN)
• 作者: Philip A. Adey;S. Akçay;M. Bordewich;T. Breckon

Region Based Anomaly Detection with Real-Time Training and Analysis

• DOI: 10.1109/icmla.2019.00092
• 发表时间: 2019-12
• 期刊: 2019 18th IEEE International Conference On Machine Learning And Applications (ICMLA)
• 作者: Philip A. Adey;Oliver K. Hamilton;M. Bordewich;T. Breckon