local-binary-convolution I create a version of caffe. step1: add a new method of weight initialization called "bernoulli distribution" in filler.hpp. step2: set the learning rate of convolution layer as zero so that the weights will be not updated.