Fit Without Fear: An Over-Fitting Perspective on Modern Deep and Shallow Learning

[@redberries8039]

'All local minima are global?' ..how about lots of local minima fall below the threshold needed for 100% accuracy ... they are not all global minima, rather only 1 will be?

[@vath85817]

Consider the function y = |sin(x)|. Lots of local minima, which are all global too. As you said, using accuracy as the loss func doesn't provide enough precision.