当前位置 :

新闻中心

关于佐治亚理工学院Chin-Hui Lee教授学术报告的通知

  讲座题目:

  Decision-Feedback Learning from Big Data: A Paradigm Shift from Density Approximation in Pattern Recognition

  主讲嘉宾:Chin-Hui Lee

  时 间:2013年12月16日下午15:00-16:30

  地 点:科大西区科技实验西楼一楼117会议室

  主办单位:语音及语言信息处理国家工程实验室

  报告摘要:

  Recently discriminative training (DT) has attracted new attentions in speech, language and multimedia processing because of its ability to achieve better performance and enhanced robustness in pattern recognition than conventional model training algorithms. When probabilistic densities are used to characterize class representations, optimization criteria, such as minimum mean squared error (MMSE), maximum likelihood (ML), maximum a posteriori (MAP), or maximum entropy (ME), are often adopted to estimate the parameters of the competing distributions. However the objective in pattern recognition or verification is usually different from density approximation. On the other hand decision-feedback learning (DFL) adjusts these parameters according to the decision made with the current set of estimated discriminants such that it often implies learning decision boundaries. In essence DLF attempts to jointly estimate all the parameters of the competing discriminants all together to meet the performance requirements of a specific problem setting. This provides a new perspective in the recent push of Big Data initiatives especially in cases when the underlying distributions of the data are not completely known.

  The key to DFL-based DT is that a decision function that determines the performance for a given training set is smoothly embedded in the objective functions so that their parameters can be learned by adjusting their current values to optimize the desired evaluation metrics in a direction guided by the feedback obtained from the current set of decision parameters. Some popular performance criteria include minimum classification error (MCE), minimum verification error (MVE), maximal figure-of-merit (MFoM), maximum average precision (MAP), and minimum area under the receiver operating characteristic curve (MAUC).

  In theory the DFL-based algorithms asymptotically achieve the best performance almost surely for a given training set with their corresponding features, classifiers and verifiers without using the knowledge of the underlying competing distributions. In practice DFL offers a date-centric learning perspective and reduces the error rates by as much as 40% in many pattern recognition and verification problems, such as automatic speech recognition, speaker recognition, utterance verification, spoken language recognition, text categorization, and automatic image annotation, without the need to change the system architectures.

  嘉宾简介:

  Chin-Hui Lee is a professor at School of Electrical and Computer Engineering, Georgia Institute of Technology. His research interests include multimedia communication, multimedia signal and information processing, speech and speaker recognition, speech and language modeling, spoken dialogue processing, adaptive and discriminative learning, biometric authentication, and information retrieval.

  He is a member of the IEEE Signal Processing Society (SPS) and International Speech Communication Association (ISCA). In 1991-1995, he was an associate editor for the IEEE Transactions on Signal Processing and Transactions on Speech and Audio Processing. During the same period, he served as a member of the ARPA Spoken Language Coordination Committee. In 1995-1998 he was a member of the Speech Processing Technical Committee and later became the chairman from 1997 to 1998. In 1996, he helped promote the SPS Multimedia Signal Processing Technical Committee in which he is a founding member. Dr. Lee is a Fellow of the IEEE and a Fellow of ISCA. He has published over 400 papers and 30 patents.