search

ijindal / Noisy_Dropout_regularization

11 stars 1 forks source link

question regarding the noisy label generation #3

Open devraj89 opened 5 years ago

devraj89 commented 5 years ago

Hi Ishan

Thanks for your impressive work.

Thanks for closing my other question. Now I could understand your code. I do however have a question

In line https://github.com/ijindal/Noisy_Dropout_regularization/blob/master/Random_noise.m#L28 it is written that we are generating a sample from the distribution vect_prob = mnrnd(1,w(tt,:)); for the tt class

This has been done for both the files Uniform_noise.m and Random_noise.m.

According to your paper, the matrix is column stochastic and hence we should sample from each column. So should not be the line replaced with vect_prob = mnrnd(1,w(:,tt)); for the tt class ?

I mean for the Uniform_noise.m it should not matter but for the Random_noise.m it should matter.

Please let me know this or is my understanding wrong?

Thanks Devraj