It would be nice to have (quantile regression)[http://en.wikipedia.org/wiki/Quantile_regression] in pylearn2.
It can be implemented as a last linear layer of an MLP with a specific cost.

One use case is to produce an estimation interval rather than a single regression value.