It would be really nice if it was like that, and I guess they also hoped that they would get such a result "unexpectedly."

I wonder if there was a technique to regulate the training so their higher layer will be similar despite the input are fundamentally different.