Improving validation and interpretation of DL.

For all the leaps forward in NN models, I have a feeling it's at risk because of a lack of trust. It's potential may be limited if the public perceives it to be too brittle or maladaptive.

DARPA is currently trying to work on DL models that are truly interpretable. If that gets sufficient progress I think the public/regulators will be much more willing to let software control aspects of their lives than a black box