http:/parnec.nuaa.edu.cn Handwritten digit recognition aided by meta-features Outline. The proposed method of incorporating meta-features to"standard" learning can be summa rized as follows 1. Define simple features and meta-features(Fig. 3) Meta-features should be informative on feature relevance or relations between fea 2. Select a subset of features by their meta-features 3. Define a gaussian process prior of meta-features to weights mapping, i.e. covariance function C(u4,u)(eg.7) 4. Build the covariance matrix C and solve svm by kernel K(xi, xi)= x Cx, or linear transforma- on Xi=C/x:(eq. 4, 5). Calculate result weights (w) and assign w= C1/2w. Only w is required for the classifier(test time) 5.Optimize selected features by tuning value of meta-features(e.g. using cross validation). The concept of feature selection by their meta-features appears in Krupka et al. 2006Company name www.themegallery.com Handwritten digit recognition aided by meta-features