AI - part 4 - Collaboration
In the penultimate chapter of his 5-part blog post on artificial intelligence, Richard Develyn, CloudTrade CTO, emphasizes the importance of hybrid problem-solving applications where human intelligence and AI functionality are blended.
“Purity” in the search for artificial intelligence can be a little bit of a distraction. I’m not questioning that, to some people, there is a great deal of importance attached to its ethical, or perhaps even spiritual, dimensions. It’s just that, when you finally finish getting emotional about proving or disproving whether a machine can be intelligent, it does allow us to get on with the business of solving real-world problems whose solutions appear to require intelligent (whatever that might mean) processing, using the most effective means at our disposal, whether that means intelligence from us, or from a computer, or from neither or from both.
Most of the advances made in intelligent problem solving come from an application of human and machine intelligence working together. Unfortunately, machine intelligence “purists” tend to gloss over the human contribution when they report their results or proselytise their achievements.
Facial recognition is a case in point. Computers are now as good as humans at recognising faces, but that’s not just down to advances in neural network technology: most of the credit should be placed at the door of all of those scientists and engineers, all over the world, who continue to constantly improve and refine the biometric algorithms which make the technology possible.
The “trick” to figuring out how to tell one face from another was not discovered by neural networks. That was human intelligence at work. Neural networks solve the much simpler problem of finding correlations between a relatively small set of measurements which human researchers have determined form the “signature” of your face. Without such a signature, facial recognition would be no better than image recognition, which is nothing like as impressive.
(My advice: don’t ever go on a blind date with a neural network. If it knows who you are, it can pick you out in a crowd, but if it has to recognise you by, say, the carnation in your lapel or the fact that you’re carrying a copy of The Times, then it’s just as likely to start chatting up a newspaper stand or a potted plant in the corner.)
I suspect that the reason that humans are so good at facial recognition (i.e., we don’t need biometrics), as well as image recognition, is because (a) our neural networks are at least two orders of magnitude bigger, (b) we use scope reduction and bias and (c) we employ a lot of world-view context when we try to figure out who somebody is. These last two points show up in our own human fallibilities: we tend to be better at recognising people from our own culture, and sometimes we can be fooled into thinking we’ve seen someone who turns out to have been someone else because of, say, the clothes they were wearing, or the way they were walking, or whether they were using certain characteristic gestures, or even if they were simply present in a particular time and place that we always associate with them.
The fact is that when we recognise people, we don’t just look at their faces. This may turn out to be one of the most profound statements that we can ever make about the difference between human intelligence and machine intelligence. This organic neural network which we have inside our human brain is not only multi-purpose, it’s constantly solving multiple problems simultaneously, especially if they’re related, and feeding the findings of one into the input of another. Oh, and reprogramming itself at the same time. Our neural networks don’t just learn “faces”, they “live”, and in the process of living pick up not only faces but everything else that is going on around them.
This gestalt approach is what makes our intelligence so mysterious, so impossible to analyse, and so irreproducible. It probably also explains why I started out in this blog post trying to get away from the romantic notions of trying to prove or disprove machine intelligence but ended up doing almost exactly that – from a disproving perspective!
I can’t imagine that anyone working on facial recognition biometrics has any idea about how their brains are assimilating the problem and then leading them to their solutions. This is genius at work. The results of their thinking is then fed into a machine-intelligence engine, probably based on a combination of deterministic or “reasoning” style AI with statistical or “intuitive” style AI, in order to generate the final product. This hybrid approach shows the first of two ways in which human and machine intelligence can work together, in this case the “human” kind being an enabler to the “machine” kind. When it happens the other way, what is known “collaborative” AI takes place, which will be the subject of my next blog post on Machine Intelligence.
Whichever way round it is, however, the trick seems to be to make human and machine intelligence work together, rather than try to make the latter take over all of the work of the former. That’s how we’ll collaboratively take over the world