Recently, I started learning GCN managed to run examples codes.
I konw that in forward propagation stage, the hidden features will be computed with learnable parameters W and aggregated with neighbors' feature. And parameters W will be updated in backward propagation stage.
But I wondered that will the hidden features be updated in backward propagation stage? I try to search on the Google to find out , but got nothing.
So, could someone tell me whether the hidden features will be updated in backward propagation stage?
I am wondering about why "bachpropagation will change the hidden features"?
That should not be happened. BP will only update the weight matrix you defined in your layer.
Hello,
thank,you!