0
$\begingroup$

I have for example simple neural network - one neuron and 2 variables. For example we have binary AND action(0 AND 0 = 0 AND 1 = 1 AND 0 = 0, 1 AND 1 = 1). So we need to find such weights of w1*x + w2*y = T( or y = k*x + T, where k = -w1/w2) to neural network can work. Graphically it look's like:
enter image description here

So as you can see T - is not any random value. If T = 0.5 there are no such w1 and w2 to make it work (you can't draw line y = kx + T to separate 1;1 from other points). To learn my neural network I use simple learning with teacher.

$\endgroup$
  • $\begingroup$ I think the idea that a neutral network is finding a line that separates the classes is somewhat figurative. That information is implicit in the learned weights; the neural network doesn't literally output a slope & intercept. $\endgroup$ – gung - Reinstate Monica Dec 13 '15 at 20:05
  • $\begingroup$ Yes, I understand that line is just a figurative graphical model, but it show what I mean: How can I choose border value? In some cases random border value lead to unsolvable system. Or I just need to use more neurones? $\endgroup$ – Rustam Salakhutdinov Dec 13 '15 at 20:22
0
$\begingroup$

I found an right common decision: T must be random value and it's must to change in time of learning. For example:

T += error*learningRate
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.