SNGP | Laplace RF Precision update inconsistent with the likelihood

[nikhil-dce]

For the CIFAR10 example, the precision matrix update (in the LaplaceRandomFeatureCovariance class in Edward) is based on gaussian likelihood, whereas in the loss function cross-entropy is being used.

The batch precision update seems different than in the SNGP paper. The update should be based on equation 9 of the paper. Is this observation correct or am I missing something?

#question

[dustinvtran]

hey @nikhil-dce! were you able to get this resolved? ccing @jereliu

[jereliu]

Hi @nikhil-dce !

Yes you are correct the released implementation in UB used Gaussian likelihood - the main reason is to keep code simple, and performance-wise we did not observe a significant difference in ECE (and it seems to hurt OOD a bit).

[nikhil-dce]

Hi @jereliu - Thank you for getting back to me and answering my question. It might be nice to add a comment somewhere in the paper/implementation regarding this deviation from the paper. Thanks again!

Hi @nikhil-dce !

Yes you are correct the released implementation in UB used Gaussian likelihood - the main reason is to keep code simple, and performance-wise we did not observe a significant difference in ECE (and it seems to hurt OOD a bit).

[jereliu]

Thanks Nikhil! Yes will add a note to the implementation in an upcoming code update. Will report here when the pull request merged.

Hi @jereliu - Thank you for getting back to me and answering my question. It might be nice to add a comment somewhere in the paper/implementation regarding this deviation from the paper. Thanks again!

Hi @nikhil-dce !

Yes you are correct the released implementation in UB used Gaussian likelihood - the main reason is to keep code simple, and performance-wise we did not observe a significant difference in ECE (and it seems to hurt OOD a bit).