A Stochastic Approach to Classification Error Estimates in Convolutional Neural Networks