Dissertation > Excellent graduate degree dissertation topics show

Several Research Topics on Learning with Side-Information

Author: ZhouYongLei
Tutor: WangJingChun
School: Tsinghua University
Course: Control Science and Engineering
Keywords: Side-Information Sample points on the constraints The samples point on learning Invariance distance metric Manifold structure sharing
CLC: TP181
Type: Master's thesis
Year: 2005
Downloads: 91
Quote: 1
Read: Download Dissertation

Abstract

This paper mainly discusses several machine learning-based Side-Information. Is different from the concept in the field of communications, Side-Information in this article refers to a sample level constraint information for characterization of whether two samples belonging to the same class. Obviously, Side-Information is more general than the sample labels supervision information. Therefore, to study how the introduction of the learning algorithm Side-Information will have a very widespread theoretical and practical significance. We recognize that: (1) in a non-supervised learning, Side-Information can guide the algorithm to get more meaningful results; contains information about the sample changes (2) sample points on such information is useful to design classifiers Therefore, we can sample point to design a classifier based. This research work mainly includes the following aspects: First, we propose a framework for a Side-Information and a large number of unlabeled samples self-reinforcing feature extraction and analysis algorithms design iterative self-reinforced ingredients this framework ( ISERCA). This framework draws on semi-supervised learning idea: the design of the space using Gaussian mixture model and Boosting semi-supervised classification, the Side-Information in sample points; then extract the relevant characteristics of the use of existing and expanded Side-Information , the samples are projected to a new feature space, repeat the previous step. The experiments prove more robust than other similar algorithms, ISERCA and effective. Secondly, we discussed from the point of view of design invariance manifold distance using the sample points to ideas and design manifold classifier proposed adjustable nearest feature line classification algorithm (TNFL). TNFL a classifier based on a similar sample points on the design, it improved the disadvantage of the classification algorithm of the nearest feature line (NFL). The experimental results in a lot of real data sets verify the validity of TNFL. Finally, our local manifold structure shared classifier (SSLMC), based on the work of other researchers. In in SSLMC local similar sample points difference probability describe the local manifolds cut subspace. Meanwhile, in order to overcome the problems the manifold classifier performance degradation caused by the less training data, noise introduced in SSLMC \\

Related Dissertations

CLC: > Industrial Technology > Automation technology,computer technology > Automated basic theory > Artificial intelligence theory > Automated reasoning,machine learning
© 2012 www.DissertationTopic.Net Mobile