I want to vary white noise by 5% each time and I want this variation to have an impact on the output error of neural network. If we assume that z is data matrix and R is noise random matrix, so R= normrnd(mu,d*sigma, M,N) where mu is the mean and set to be zero and sigma is standrad deviation is set to 0.05 and d is varied from 0 to 5. So each time, we set d=0,1,2,3,4,5, then we add z(noisy data)=z(original data)+R (noise). Is that correct.
I think you should first get an estimate on sigma from your original data matrix. You could either calculate sigma from the whole matrix or you could calculate it for each column depending on your application. So you should add noise to your data based on a percentage of the original standard deviation of your data.
Thanks Saper. I just figured this out. You are right. First come up with the standard deviation of my data and then vary it into 5% each time.
Your computation seems to be correct. Note that d=0 corresponds to the state of noise with constant level condition. Also I want to note that a conventional model for generating noise is Gaussian (Normal) distribution. To generate a random M times N matrix with mean mu and standard deviation (d*sigma) you need to run R=mu+d*sigma*randn(M,N),. Then if you add R to the original data, you will have noisy data.
- Badges
- Science topic