>What we need is much more than a clever idea: it's a paradigm shift. There's probably some sort of way to make a system that can learn from very few examples and at a very short time, like we do ourselves. Machine learning is certainly not that way. We need to find another way.

That's more like probabilistic methods, which also fight overfitting by regularizing (with a prior distribution) and allocating probability density in a smear around the training data rather than concentrating it as hard as it can (as optimization methods usually do, so to speak). And then hierarchical Bayes is what's used in many of the neat one-shot learning demos, and in theoretical neuroscience these days.