应用子空间信息准则选择模型参数

Selecting model parameter by subspace information criterion

在监督学习中,模型的选择直接影响学习算法的推广能力.子空间信息准则是模型选择的一种新准则,它在一些假设条件下,给出推广误差的一种无偏估计.研究再生核Hilbert空间时,将子空间信息准则应用于模型参数选择问题,也证明子空间信息准则是推广误差的一种无偏估计.

In supervised learning, model selection influences the generalization capability of the learning algorithm directly. Subspace information criterion is a new criterion for model selection, it gives an unbiased estimate for the generalization error under some assumptions. Discusses the model parameters selection for the approximation function by SIC in reproducing kernel Hilbert space, and it has also been proved that the SIC is an unbiased estimate for the generalization error.