During Backward_* in the Slience layer, it performs the following operation when propagate_down[i] is set to true:

caffe_set(bottom[i]->count(), Dtype(0), bottom[i]->mutable_cpu_data());

and similarly for the GPU.

Usually this code will not run because the Silence layer does not generally need backprop. However, if you force the network to do backprop, this will have the consequence of overwriting the bottom data blob with all zeros.

I think the correct behavior is to set the bottom diff to 0:

caffe_set(bottom[i]->count(), Dtype(0), bottom[i]->mutable_cpu_diff());

Should this be changed?

Thanks!