Incremental Learning

As @albanD explained in the other topic you would have to zero out the gradients of your fixed parameters.