False positives with no discernible signal about aifororcas-livesystem HOT 3 OPEN

Per Teams conversation today, confirmation from Prakruti and Dave that:

1. The last model re-training was in 2020 (not 2021)
2. The model that is currently deployed is the same as what was deployed before migration between Azure subscriptions.

Dave: "PowerPoint refers to retraining on 11-16-20 that reduced false positives by 70%"

Prakruti: "No updates made since 11/15/20 as michelle_mit mentioned that is the latest one. Nothing retrained in 2021. This is the same model used with the old subscription also.

This regression is because the samples are out of domain for our model for whatever reason not because there was any retraining done. If we retrain with these neg samples - performance will improve."
... "As the date suggests that retraining was performed in 2020 not in 2021. Both subscriptions have the same model which is why I think we can eliminate redeployment in 2021 as the cause of the regression."

from aifororcas-livesystem.

Example today of a candidate with high average confidence (>75%) and many false positives, but no SRKW or other obvious biological signals, just prominent "gurgling" water sounds with low frequency modulations that sometimes mimic humpback non-song vocalizations --

from aifororcas-livesystem.

I wonder if it would be useful to train a general model then fine tune for individual hydrophones.

from aifororcas-livesystem.