That's why I'm still skeptical about whether we are heading the right direction with current DNN techniques. We're basically brute forcing extremely complex statistical models that rely on countless data to build those regressions because we don't yet know a good model for training with minimal data.