"Deep learning is fantastic at for instance building appearance models and exploiting hierarchical nature of natural images..." Agreed. And contrast this succ

"Deep learning is fantastic at for instance building appearance models and exploiting hierarchical nature of natural images..." Agreed. And contrast this success with the lack of success of first-principles latent-variable modeling for natural images. A lot of very good researchers spent decades building multi-layer probabilistic models for natural image structures - I'm thinking about the Grenander school, for example. The jury is still out (I think) on the ultimate value of that approach. But for classification, it turns out to be much more tractable to use DL. You don't need all the semantic information the multi-layer model contains to tell a car from a truck. As you say, it's better to view these approaches as complementary.