[Work Log]
- Implementing covariance functions
- Implement generalized marginal likelihood
- Implement fast generalized marginal likelihood
Implement generalized marginal likelihood
Rewriting curve_ml3.m
Based on curve_ml.m
Kernel now takes two sets of indices, curve-index and view-index.
Kernel is now passed-in as a lambda.
Need to decide whether to use matrix inversion lemma for global offset and linear prior. Currently being used; not using would be simpler; try both and compare results.
Testing new marginal likelihood
Creating new version of test/test_ml_end_to_end.m, named test/test_ml_end_to_end2.m. This version uses the new likelihood format, and compares against the old.
New curve_likelihood
Need to write new version of curve_likelihood.m to handle new likelihood format.
It's going to be difficult to compare to older version, since we've changed the correspondence method...
I'll have to add a flag that simulates the old correspondence method, but stores it in the new likelihood format...
01:33:52 PM
I decided to write a separate function for the old correspondence method, called corr_to_likelihood_legacy.m. Moved clean_correspondence3.m to corr_to_likelihood.m.
The legacy function isn't exactly identical to the old function. In particular, the index values might be slightly different, since it depends on the old function's global mean curve, which we don't compute in the new function. If it's important, we can call the old function, and then call the new one. This is a debugging function, so speed doesn't matter. For the moment, we'll accept the small difference, and use this function simply to confirm that we're in the ballpark.
03:06:24 PM Test is implemented. Troubleshooting syntax errors...
04:11:17 PM
Elusive NaN error in corr_to_likelihood_legacy.m.