Low variance and high error

[advaitb]

Hi @ziyiyin97,

I'm getting a few results where there is little to no variance in the posterior and high error (in some regions) and vice versa during testing. Any reason this might be happening? Is it likely overfitting? This is using WISE.

[ziyiyin97]

Thanks for sharing these interesting results! I think this reflects the "inaccuracy" of WISE in predicting velocity models in this particular example. This can be due to many possible reasons related to the existence of the "amortization gap". Some reasons that come to my mind and their corresponding "further investigation steps" include

• network not trained well enough (can check the objective function)
• training samples are limited (can check some training samples instead of unseen test samples)
• you can also try to see some posterior samples and see what they look like

A lot of these could potentially be addressed by WISER