I keep hearing that neural net training is non-convex. The objective function is non-convex.
My question is:
What is the proof that the learning process is non-convex? How can you prove it is non-convex for say max likelihood training?
Is training wrt all objective functions non convex?
I am unable to find answers to these questions as most online resources never care to explain the non-convexity and take it for granted.