“Online/continual learning and data efficiency are two of the most plausible remaining limitations of the current deep learning paradigm.”