The quest for ‘I’ in Artificial Intelligence has been refocused with the rise of Artificial General Intelligence (AGI) as a branch of AI focused on intellectual tasks, as opposed to advances in computer vision, simplistic transactional conversations and games. This is not to say the advances in these areas have not been impressive and important as they certainly have.
The point is that getting machines to perform heretofore impossible tasks that are taken for granted by humans and even animals is not getting us any closer to human intellectual capabilities, which is the original and true goal.
The recent announcement of AlphaGo Zero team, one of the leading ones in all of the contemporary AI, demonstrates not only that machines can outperform even the best humans in strategic games such as Go, but also that such feats can be accomplished in unsupervised fashion with self-learning, instead of enormous amounts of labeled data for supervised learning.
This seeming aversion toward supervised learning in A(G)I as actually understandable, as lack of quality data has been precluding AI advances for ages. But one should be very careful in trying to generalize impressive results in narrow specialized fields such as strategic games to great progress towards AGI.
One of the key questions is, and has always been, the importance of context and previous knowledge and experience. Isaac Newton famously said – ‘even if I was able to see further than the others, it was by standing upon the shoulders of giants’ where he explicitly acknowledged tremendous importance of previous knowledge. Focusing on raw smarts and intelligence alone is grossly misguided.
Are we to seriously believe that a superior Go player would also be miraculously the best mathematician, composer, politician, writer or another superior intellect just from unsupervised learning? Of course not. Archimedes belongs to the pantheon of the smartest people ever together with Newton, but he did not know calculus. With his great mind he saw glimpses of it by calculating the surface under parabola but it took hundreds of years for Newton and Leibniz to come up with the general principles of calculus. Xeno is considered a great philosopher but his paradoxes are now simple exercises at Calculus 101 level.
These, and endless other examples, show that superior intelligence at even at the highest human level comes only in the context of bodies of previous accumulated knowledge of our civilization. Attempts to achieve AGI by unsupervised learning alone will be doomed to failure as they will be unable to leverage all the we have learned so far. This is not to say that unsupervised learning will not play a role. It certainly will, for instance by being applied to mountains of said previous knowledge.
The point is it will not be alone, and that it will be coupled with supervised learning that leverages these tremendous bodies of knowledge and experience of our civilization. And this huge treasure is not only large in size, but also tremendously rich in terms of structure, that of the Reasoning Graph. This is the focus of our work and a topic for a future post.