Some questions from a person who's never done anything to do with machine learning: -Is the idea to pick features that most define your data set? For example t

Some questions from a person who's never done anything to do with machine learning: -Is the idea to pick features that most define your data set? For example the spiral is sort of periodic in x1 and x2 directions so I picked the sine functions as features to try and predict it -Is it best to have the max amount of hidden layers and max neurons? (I noticed it seems to oscillate on the output when I add a lot) -What exactly are the graphs on the hidden layer neurons? To me it looks like at each node we're combining the graphs of previous nodes but with different weights for each? -Does each node represent a sort of intermediary prediction of the original dataset? -If someone were to code one of these neural nets, would they manually code the properties of the hidden nodes? -How is this thing actually learning? Is it just some complicated error minimizing algorithm?