truongthinhs at January 3rd, 2008 08:21 — #1
I am freshman in NN. Now, I am working Back propagation. I have the analogue input and binary output. I dont know what activation functions I use for backpropagetion with the single hidden layer.
Please, help me.
I program the backpropgation based on errors. If I used sigmoid function then result is not good, and step function I cant.
Thanks for help.
reedbeta at January 3rd, 2008 13:43 — #2
What do you mean by "result is not good"? Sigmoid and similar functions are the standard choice for multilayer feed-forward networks.
nyx at January 3rd, 2008 13:59 — #3
You can use the sigmoid function... But you should remember that it's output is asymptotic... The sigmoid function "goes to 0" at minus infinity, and "goes to 1" at infinity, but never actually reaches those values. Training your network to reach them can result in very large weights, which does not help convergence.
If you use the sigmoid function, you should probably train your network to output values in a restricted range like [0.25, 0.75], the lowest representing 0, and then interpret outputs under 0.5 as being 0, and over 0.5 as 1.
truongthinhs at January 4th, 2008 01:27 — #4
Thank Nyx and Reedbeta very much!
I used the sigmoid functions for my problem. But the output signals are only in range (0.49 - 0.50). I feel they are not good. Can I interpret outputs \< 0.5 like 0, > 0.5 like 1? If I do that, the error of problem is large.
Can You help me!