Does it make sense to optimize Gaussian Process hyperparameters during active learning?

[tumble-weed]

Hi,

In the active learning for regression example, we have used gaussian processes. While the sklearn version seems to keep its length scale and noise parameters static ( maybe i am doing something wrong), other implementations allow for modifying these using gradient descent ( for e.g. gpytorch).

Under batch learning circumstances we would have the whole train set , and tuning the hyperparameters to maximize log likelihood makes sense, but does it also make sense to do so while performing active learning, and having datasets of really small sizes?

[cosmic-cortex]

Hi!

I have absolutely zero knowledge regarding this question :) I don't want to state something which is false so I rather not give any advice here. Instead, I'll keep this issue open, hoping that someone will know the answer.

[tumble-weed]

ok thanks.

On Mon, Nov 9, 2020 at 12:16 PM Tivadar Danka [email protected] wrote:

Hi!

I have absolutely zero knowledge regarding this question :) I don't want to state something which is false so I rather not give any advice here. Instead, I'll keep this issue open, hoping that someone will know the answer.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/modAL-python/modAL/issues/107#issuecomment-723798942, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABR2NXP6F7H7AGX53USWGXTSO6F4LANCNFSM4TG4F4ZA .