Does it work?

[gillonba]

The theory behind this project is interesting and I have been wondering if it would be worth trying in my own projects, but I didn't want to waste my time if it doesn't work. Do you end up building a model that can detect the correct class of objects beyond the scope of the original training data or do you end up with a ton of false positives as the system has no way of self-correcting? Does it end up going the opposite direction and overfitting? I would be interested in hearing how it worked out for you