For the first step of a project I've just begun, I need to process huge amounts of unlabeled 3D data to figure out which parts are buildings (houses, etc.), and which aren't. I've got a number of heuristics for filtering out noise, but ultimately, since this is art, it's going to need a human touch at first. As such, I figured I might use a supervised learning system to teach a neural net what is and isn't a building. However, I'm not sure of the best way to do this. I want the software to tell me it's a building if it is one and tell me it isn't if it isn't, but I also want to know when it's not certain, so that I can advise it. From what little I know of Bayesian nets, it seems they may be appropriate for trying to determine cause and effect between 3D geometry and "building-ness" while still maintaining uncertainty, but I really don't know much about them. Might it be better just to have a reinforcement learning system that gets positive reinforcement for a correct answer, high negative reinforcement for an incorrect answer, and slight negative reinforcement for asking? Or are there any other approaches I haven't considered yet? Thanks, Josh