In the example I gave about anomaly detection, what’s being detected as normal or an anomaly doesn’t need to be labeled by a human. This just comes down to probability of temporal data. This temporal data can come from segmentation which again doesn’t need human readable labeling and can be produced from an automated process. You can automate such a system to alert someone if something is an anomaly.
Edit:
Also want to add there are more traditional avenues of which non-labeled data can be used.
Unsupervised Learning where you read in arbitrary data and cluster for some sort of probability based system.
Another good thing to look at is a system with “some” labeling like Semi-supervised Learning.
Any situation where you can do binary classification doesn’t need any true labeling as long as you can evaluate attributes of some object, complex or not.
Edit: Also want to add there are more traditional avenues of which non-labeled data can be used.
Unsupervised Learning where you read in arbitrary data and cluster for some sort of probability based system.
Another good thing to look at is a system with “some” labeling like Semi-supervised Learning.
Any situation where you can do binary classification doesn’t need any true labeling as long as you can evaluate attributes of some object, complex or not.