Multiclass classification has one too many latent GPs? #1903
Replies: 2 comments
-
Yes, technically with C latent GPs it's an overparametrized model, you can of course implement a likelihood yourself that just requires C-1 latent GPs (e.g. by simply adding a constant zero output as you suggest). I don't know if that necessarily makes for a better model though. If one of your classes is "don't know", then simply fixing the latent output for that to 0 might be fine. If all your classes are 'equally valid', then you might end up with a more easily interpretable model when you have one latent GP per class. Which one is easier to optimize might depend on the data. In any case, I would recommend Softmax likelihood (with stochastic optimization) over Robustmax, it seems to work better in practice. |
Beta Was this translation helpful? Give feedback.
-
(It's not a bug, it's a modelling choice.) |
Beta Was this translation helpful? Give feedback.
-
In the multiclass classification model, using a robustmax likelihood, there are as many latent GPs (num_latent) as there are classes (C).
https://gpflow.github.io/GPflow/2.5.1/notebooks/advanced/multiclass_classification.html
When one of the latent GPs is the largest, then that class is predicted.
Shouldn't num_latent be equal to C-1 however? As one of the GPs can be set to 0 (or any other constant) instead, without loss of generalisability.
Is this not correct, and if it is correct, is there any easy way to implement this?
Thanks,
Evan
Beta Was this translation helpful? Give feedback.
All reactions