In this note Justin Domke says that
In practice, neural networks seem to usually find a reasonable solution when the number of layers is not too large, but find poor solutions when using more than, say, 2 hidden layers.
But in Bengio's remark, he says
Very simple. Just keep adding layers until the test error does not improve anymore.
There seems to be a conflict. Can anyone explain why they suggest differently? Or am I missing something?