[Work Log] 2

Experimenting with ou_perturb_kernel.m.

Was getting weird results, where the ML approached infinity as noise_variance approached zero.

Realized that the set of precision matrices needs to be updated every time the noise sigma changes.

This is an oversight that has tripped me up before. Need to try to avoid it in the future.

It is possible (likely?) that this transformation is as simple as multiplying the precision matrices by \(\sigma_n* / \sigma_n\). This would avoid a semi-expensive Hessian calculation for each point, which could be a bottleneck during training.

TODO

See previous entry