How to find the largest errors on images when using SVM for image classification

Question

I\'m writing an image recognition pipe using HOG and SVM. I have 10 different classes and I want to find for each class the image that has the largest error. I\'m using skle