What are the benefits of a deep neural network architecture over a shallow architecture with many computational elements.

What are pros and cons of these two architectures and its limitations?

  • $\begingroup$ Each layer of deep is about finding unique (transform invariant) features starting on very small scopes and then increasing as depth increases. This is typically accomplished by the combination of convolutional neural networks and restricted boltzman learning between layers. There are alternate formulations that achieve similar results, but these are the "textbook" deep belief network components. In typical feedback cases there is not the squential abstraction of features to larger and larger scopes. Alpha-go and other "super-performers" used model-adaptive model-predictive control w/ nns. $\endgroup$ – EngrStudent - Reinstate Monica Aug 6 '16 at 18:45