View in #general on Slack
@Ashish: In google course section on online learning it says - “Needs monitoring, model rollback & data quarantine capabilities”. Does anyone know what data quarantine mean in this context?
@vlomonaco: I guess the ability to isolate outliers and remove them from the training data? Not sure though, just guessing!
What is this course exactly?
@Gabriele_Graffieti: I’m just guessing but maybe the ability of maintain data separated and not aggregate it to a single point? As an example sensor data is not sent to a central server but keep in a local database and never or rarely moved.
Let us know the context and if you find out what are data quarantine capabilities
@vlomonaco: for sure it’s not a nomenclature I’ve found somewhere else in CL…
@Ashish: I found this in a microsoft doc: How do you recover from adversarial contamination of your training data?
-Can you isolate/quarantine adversarial content and re-train impacted models?
-Can you roll back/recover to a model of a prior version for re-training?
This doc is not for online training, but it makes sense in terms of online training.
as Vlomonaco mentioned it applies to outliers as well. Any data we do not want model to see. What would be some ways to do it in online learning scenario where training data is streaming in?