Artificial Intelligence (AI) has come a long way since Alan Turing proposed the Turing test in 1950. In last few years, the field of AI has taken some spectacular strides forward, from IBM’s Deep Blue computer defeating Gary Kasparov in 1997 to the myriad autonomous cars currently being tested.
The major advantages of Deep Blue lay in its memory and computation powers. It could compute faster and thereby it could analyse potential positions more accurately than a human. But this computer could not learn anything by itself. It just performed according to the inputs given to it.
Ability to learn is one of the most important facets of human intelligence. In 2016, AlphaGo, a Go-playing computer system defeated the world’s best Go player, Lee Sedol. With machines mastering Go, a 2,500-year-old board game with more possible moves than there are atoms in the universe, it is now clear that machines can learn, and that too possibly better than a human. With this epoch-making move, can we make the claim that AI has peaked?
A few years ago, Prof. Hans Moravec of Carnegie Mellon University developed a “landscape of human competence” map that represents human activities in terms of the difficulty for computers to emulate. The lowlands in this map are activities like rote memorization and doing arithmetic calculations. Today, even a pocket calculator can do these activities far better than humans.
The next higher peaks in Moravec’s map are activities like playing chess, which the computers are already doing better than humans. Even higher peaks are activities such as speech recognition and language translation. Computers are already doing a fairly good job doing these activities too. It is just a matter of a few years and computers will achieve a perfection in these activities that is far superior to what humans can do.
All these activities that computers have done an excellent job of surpassing human competence have certain common properties. All these activities are repeatable behaviours that have got a predictable pattern. All these activities are clearly defined, based on preset rules and regulations. These rules are consistent on a temporal scale. The inputs into these activities are stable and the factors that impact the output are fixed and accurately measurable. Besides all these advantages, there is abundant source of data to study any of these activities over millions of instances.
The next level of activities on Moravec’s map are very different from activities that machines so far have done an excellent job of mastering. The next level activity AI hopes to master is social interactions. The ultimate peaks in the Moravec’s map are activities like book writing and arts. These activities involve creativity, totally unpredictable patterns of human brain activity.
From now on the complexity of mastering human activities take on a very different level of difficulty. So it calls for a new level of expertise and competence too. From now on, AI is entering the complex world of human behaviour. There are very few clear rules that govern human behaviour. Even if there are, those rules vary with even small changes in the context. Even in the same context, different people will behave differently. More importantly, emotions are involved in all facets of human behaviour. Unless AI understands and incorporates emotions, it will always remain artificial.
Human behaviour is the result of electro-chemical reactions involving billion of neurons, trillions of synapses, several neurotransmitters. Even the most optimistic neuroscientist will admit that we know very little about the functioning of the human brain. Every day new discoveries are adding to our understanding of the human brain.
Neuroscientists for years have been trying to map all the connections between the 20,000 neurons of a simple sea slug called Aplysia californica. We still haven’t managed to do that. If so, how long will it take to create an artificial connectome, a replica of the human brain involving billions of neurons of varying characters? What makes mastering human behaviour tougher is that a vast majority of brain activities occur at a non-conscious level. So even an individual does not have an understanding of his own behaviour. How can AI replicate behaviour that a person himself is not fully aware of?
To understand social interactions it is not enough that one understands the functioning of an individual brain. One also needs to study the synergistic reactions when several individuals are involved in group behaviour. So, the next stage of AI, mastering the process of social interactions, is going to be a very tough one to cross.
Human behaviour is multidimensional in nature. So to fully inculcate all the intricacies of human behaviour into the machines, AI has to become a multidisciplinary science. Learning has to be taken not just from neuroscience but also from the world of other human behaviour sciences such as evolutionary biology, sociology, and anthropology.
Machines have shown that once they learn to do a cognitive task, they are able to do it far better than a human being. So the learning of new cognitive tasks by machines is a welcome development. But machines are still a long way from competing with humans in all cognitive tasks. For AI to become less and less artificial and become more human, it has to retool itself. AI has to become far more multidisciplinary in its learning and approach.
Biju Dominic is the chief executive officer of Final Mile Consulting, a behaviour architecture firm.
Comments are welcome at views@livemint.com