I think any monumental leap forward is going to come from changing the way networks are motivated. Simple back propagation attempting to reduce the error from a fixed objective have proven to be useful, but doesn't actually resemble how more generalizable intelligence works.

I think getting the feedback loop integrated with something that behaves more like dopamine/serotonin/pain feedback is going to be the likely direction we'd need to go. Basically, the network needs to be able to form new objectives and recognize when it's meeting or failing at those objectives, rather than just optimizing its network to be less and less bad at predicting specific outputs.