I wouldn't claim the Neural nets don't work. We've seen many use cases where they do work (And I have looked at deep learning just a little, it may superior in ways but it doesn't seem in any way fundamentally different from the other stuff).
I would add that Support Vector Machines also work and they are similar and have a much clearer math to them [1]. But SVM and neural nets are ultimately just linear matchers on a nonlinear pattern space, they ultimately involve adhoc choices that experts learn over time.
As I said above, once you learn the maths (linear algebra, statistics, functional-analysis or whatnot), it become less basic understand and more "understanding how", a series of tweaks that experts "with a feel for this stuff" do. But this "feel" level understanding seems exactly what stands in the way of serious, rational progress on the topic.
I would add that Support Vector Machines also work and they are similar and have a much clearer math to them [1]. But SVM and neural nets are ultimately just linear matchers on a nonlinear pattern space, they ultimately involve adhoc choices that experts learn over time.
As I said above, once you learn the maths (linear algebra, statistics, functional-analysis or whatnot), it become less basic understand and more "understanding how", a series of tweaks that experts "with a feel for this stuff" do. But this "feel" level understanding seems exactly what stands in the way of serious, rational progress on the topic.
[1]http://en.wikipedia.org/wiki/Support_vector_machine