[Work Log] Likelihood Training
October 19, 2013
Training likelihood
Fitting a three-component GMM resulted in a null model having a higher likelihood than the ground truth model.
Spent a lot of time troubleshooting this and finally realized I had scaled all values up by 100x so Matlab wouldn't choke, but I wasn't accounting for this when evaluating. Blame it on this blasted cold, killing my focus!
Ground truth model looks much better than null model now.
--
Trying different blurring levels.
- 1.0 is to small.
- 2.0 seems okay
- 5.0 gives a wide variance for true-positives.
In the end, running max likleihood on this will be best, but this gives us a ballpark.
--
Trying to get KJB EM algorithm working on my 2M element dataset. Getting NaN errors at the moment.
Posted by Kyle Simek