The most successful classifier GMM-UBM uses log-likelihood estimation of training data for a given model or latent parameters like mean and variance. The drawback of using log is, it generates a 0 value if a Gaussian distribution returns unit value and produce low variance with a high change of values. The distribution with 0 values produces NaN values. In a result, a GMM model becomes completely useless to compare with the testing data. Because, this model also returns NaN values for the testing sample's likelihood measurement. We can selectively delete the distributions with the NaN values. This deletion makes the mixture distribution random for individual class. Have there any way to minimise or fix this NaN problem in the GMM-UBM when running with EM algorithm?
Dear Atiqul,
I'm not familiar with GMM-UBM, so I'm not sure if my answer has any relevance, but maybe...
In several cases, log(x) is not used alone in the source codes, but in the form of x*log(x) [e.g. this is the case for dismo::calc.deviance()]. If you find this in the source code, you can change it to
if (x == 0) 0 else x*log(x)
since x*log(x) = log(x^x), and 0^0 = 1. This will eliminate NaN-s produced by log(0).
However, please refer to other opinions: https://math.stackexchange.com/questions/357392/why-i-cant-calculate-0log0-but-can-log00
HTH,
Ákos
Using a log-likelihood estimation does not imply taking logs of data but well the log of a probability or the log of a density distribution. I don't know GMM-UBM either so you need to enter more into details for me being able to help you.
Hi,
In my opinion, you should test either the log(distribution) provides NaN or not. If yes, one can neglect this observation and not to add it in the log likelihood value.
However, doing this may lead to erroneous results, so, I propose to compute the number of NaNs and find the minimum of log(distributions), hence to add the number of NaNs multiplied by the minimum of log(distributions). This should reduce the risk to obtain erroneous optima.
Here I provide a pseudo algorithm to perform what I said.
loglikelihood=0;
counternan=0;
minlogs=0;
for t=1:n
compute f(x_{t};θ)
a=log(f(x_{t};θ));
if not isnan(a)
loglikelihood=loglikelihood+a;
if a
I faced the same problem. Plz check the data of your frames/channels. If the entire channel is zero, it results the NAN value. In this regards, you can either discard the channels with completely zero values or you can use PCA for dimension reduction.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Computer Graphics and Animation
- 3D