I’ve found exactly the opposite. In domain after domain the performance of a pure deep learning method is orders of magnitude less than that of either a traditional algorithm or a combination.

And often the CNNs are so finicky about noise or distortion that you need something as an input stage to clean up the data.