Dealing with noise in Neural Network

[sanjay]

Hi all,

The concept of noise for training a particular system for signal classification.

I am implementing a system, whereby the signals can and cannot be drowned in noise. From the network designed so far, I get excellent (well, almost excellent) results out from both training and simulation, when I have no noise to the inputs of the signals. however, when I try to check the output behaviour of the network while simulating the network with noise, My signal classification jst gets messed up. I have looked into the possibilities of overfitting and issues relating to that. But so far, haven't really encountered any success. Even trying different values for learning rate, momentum, change in the size of the network hasn't worked.

Can someone share their expertise comments regarding to the problem of adapting the neural network for signal classification with noise using method of backpropagation.

Regards

help would reallyyyyyyyyyyyyyy be appreciated

 

[me2please]

1. If NN learning converges, try several initial condition(weights).
2. I don't know what do you use as inputs. However, if NN cannot classify the noisy inputs correctly, it might be the case that the noises make the input feature spaces of differect inputs to overlap and there is not much left to discriminate. You can try feeding more data of the signal (maybe in some transform domain). It's like if you cannot tell 2 objects apart by looking at one angle, then rotate it and look at another angle.

 

[sanjay]

Hi, my neural network does CONVERGE, but the funny thing I cant figure out LOGICALLY is that why it converges for the same input set, same network parametres when I add noise of 0.5% (gaussian distribution) where as, if i train it without Noise, it doesn't converge, instead it sits in a straight line.

I have already tried different initalization functions for weights and biases, as well as different learning algorithms, at the moment, Levenberg Marquadt algorithm is doing the best job for me.

 

[newbie101]

probably because the large step when you discretize data... the small changes cant be detected... adding noise allows data to go over to next value... so no straight line

**at least thats what i read in book**