Weight update for Siamese networks in Caffe

[hjung]

By following this web-site 

http://caffe.berkeleyvision.org/gathered/examples/siamese.html

we can use a Siamese network in Caffe, which shares the weights for each layer.

But, I was wondering about how the Siamese network in Caffe updates their shared weights from the backpropagation.

To be specific, if we have 

input1 -> conv1(shared) -> output1
input2 -> conv1(shared) -> output2  ===> contrastive loss (for output1 and output2),

then, Caffe just sums up the two gradients for conv1 from the first and second networks?

Thanks for your response in advance.