-
|
Hey River community, lets say im using my model in the production for a multiclass classification problem and after a while it turned out that the model was somehow fed with wrong informations eg. wrong labels. so my question is, what should I do in this case? should I retrain the model from the scratch? is there a way to delete/extract the wrong labels from my model? Thank you! |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 3 replies
-
|
Hi @BaAL94, unfortunately there is no way to "cancel" specific learning steps a model performed. The best is to prevent such issues to happen if you can. Otherwise, one way could be to make regular snapshots of your model and revert to one of them when it happens. |
Beta Was this translation helpful? Give feedback.
-
|
Good answer. Also note that an online model implicitly forgets about past observations. So the impact of erroneous sample will fade with time. |
Beta Was this translation helpful? Give feedback.
-
|
@MaxHalford Is implicit forgetting somewhere documented? |
Beta Was this translation helpful? Give feedback.
-
|
No, not really @Robert-Schimanek. What kind of explanation would be looking for? |
Beta Was this translation helpful? Give feedback.
Hi @BaAL94, unfortunately there is no way to "cancel" specific learning steps a model performed. The best is to prevent such issues to happen if you can. Otherwise, one way could be to make regular snapshots of your model and revert to one of them when it happens.