Ib Theory And Parameter Validation Study

By compressing one variable to the "Bottleneck" variable and maximally preserving its relevance to another variable, IB theory provides a framework for extracting feature patterns hidden between variables and effectively settles the problems that many other classical feature extracting methods can't solve. Along with the embedded applications of IB theory in areas such as coding analysis, image analysis, pattern recognition, machine learning, information security and so on, the shortcomings of IB algorithms in the case of enlarging and changing of searching spaces have emerged. These problems have gain attentions from researchers on areas such as image processing, information encoding and machine learning. Some of these problems have become the bottleneck of further developments of IB theory. One of these problems can be summarized as: determining the parameter of compression variable for IB algorithms.To solve the problem of determining the compression variable parameter for sIB algorithm, this paper proposes an AsIB algorithm for automatically determining parameter based on minimum description length principle. An efficient encoding scheme is designed to estimate the description length of the solution model and the original data given the model respectively, and the minimum description length model is selected as a criterion to find the number of feature patterns hidden in dataset. Experiment results show that the encoding scheme in AsIB is efficient to recover the true feature pattern in dataset without the requirement of setting category number of feature pattern. AsIB algorithm removes the dependency of empirical knowledge for sIB algorithm, which widens its applications in areas such as automatic dimension reduction and pattern extraction,etc.