I understand that the initial weights for a NN should be kept small. So is it fine that i just select the initial weights to be random values between 0-1. I have seen some programmers use much more complex techniques (like simulated annealing) to find the initial weights. Do initial weights really affect the working of a Net, I mean one might choose to implement complicated optimizing algorithms to avoid local minima later on, and thereby train the nets to attain the best possible weights. Or is there something abt the initial weights that I haven't understood, or have missed? thanks, sidhant

Choosing small random initial weights works.

Simulated annealing can be used to try to find a good set of weights and then a gradient decent to refine the weights.