Manifold detection with false positives

If you have little detectors that only sample part of the input and fire when they detect a learnt pattern, then those detectors will have a false positive rate (eg. Bloom filters.)
Is that false positive rate a good or bad thing? If you have more than one layer of detectors then false positives probably can help you learn more sophisticated features in later layers.
You can think about it. The false positives probably are still information rich and without them each proceeding layer would get exponentially sparse until rapid nothingness.