What鈥檚 an engineer doing up to his elbows in neurons? Intelligent things, says Igor Aleksander of Imperial College London. He was one of the pioneers of neural computers that borrow heavily on the brain鈥檚 design. By the 1980s, they could learn, recognise patterns and have since found their place in everything from controlling flight simulators to focusing your camera. Now they鈥檙e going back to their roots, helping us to understand the human brain. Aleksander believes you can even talk of machines that have imagination. Jeremy Webb asked him if he was serious about the claims in his book How to Build a Mind.
That is a very provocative book title because it implies that you know how to do it. Do you?
No. But I hope to find out by doing it. When they don鈥檛 understand something engineers try to build it. But there is an intended frisson in that you might expect to be able to build a brain, but not a mind, whereas I鈥檓 arguing that a mind is an emergent property of brains one might build.
Advertisement
You call this idea understanding by synthesis. Is there another organ that we have taken apart in the same way?
The heart is probably the best example because the interplay between the neural anatomy of the heart and the modelling of the heart in a computer has been simulated on computer and understood by reverse engineering. But another fascinating example-which sadly isn鈥檛 happening in my lab-is Jack Cowan鈥檚 work at the University of Chicago. He has made a reverse model of the early parts of the visual system and then modelled the injection of drugs and predicted to a very close approximation the hallucination that the person taking those drugs would have.
Most people associate you with Magnus, the neural computer. How did that come about?
I used to be known for something called the Wisard which was built in the 1980s. It was the first neural pattern recognition machine, an intelligent machine without a mind, just a layer of neurons which took a pattern in and labelled it. It had no contemplative activity at all. In the early 1990s, we wanted to move on and develop architectures with internal contemplative activity. To do that you need networks that receive inputs and produce outputs, but in between the neurons are having a 鈥渃onversation鈥� which isn鈥檛 obvious when you look at the output. This creates the kind of inner state which human beings are enormously familiar with.
What have you done with Magnus鈥檚 inner state?
In the very early days we built a machine which was able to find its way around a real kitchen table full of cups and saucers, knives and so on. Now, remember, this is a Magnus in a virtual world and a virtual everything. First of all we would show it various objects, which it learned. But it didn鈥檛 learn to label patterns as cup, saucer and so on. It learns what these things look like. It has an internal depiction of what these things look like so when I say cup, it would visualise internally a cup.
It visualises internally?
Absolutely. It produces images on a screen. And these images tell us if it鈥檚 imagining properly or not. I might say 鈥渢hink of a cup鈥�. Then say, 鈥渢ell me where there is one of these things in the thing you鈥檙e looking at鈥�. Then Magnus uses its inner depiction of a cup to find others in its virtual world.
Are these depictions stored at one point?
No. Depictions in the brain are stored in a highly distributed fashion and that鈥檚 how they鈥檙e stored in the simulation.
Can Magnus imagine too?
Yes. We can say imagine a green triangle, even though it鈥檚 never seen a green triangle, or imagine a blue banana with red spots. And it can imagine them. We鈥檝e achieved that by using only adjectival phrases-say, blue with red spots. It doesn鈥檛 have a memory of these things but it puts features together which lets it 鈥渋magine鈥� objects never seen.
You鈥檝e changed Magnus鈥檚 name to Neural Representation Modeller. Why?
Mainly because Magnus is like Meccano. You can put together all sorts of structures. It鈥檚 probably my fault for not being clear enough, but it came across wrongly to the world. People were coming to my lab wanting to see a robot called Magnus which was totally conscious so they could have a nice conversation with him. But it鈥檚 really a prototyping tool.
And you鈥檙e using your NRM to model parts of the brain?
We鈥檙e inspired by the architectures of the brain. To take the visual system alone, it鈥檚 got 46 different areas that interact. One part of this architecture is where the visual signals come in, everything else does a lot of processing, moving things around, separating out into colours, lines and so on, and then somehow it comes back together to give us consciousness. So we start off by transferring this architecture into the NRM. Any bits that aren鈥檛 understood we evolve. We assume that there鈥檚 some sort of genetic evolution which can happen very fast on a computer that optimises the structure. But we may also end up understanding why these bits are there. We can evolve interconnections that work and then ask the neurophysiologists to tell us whether that鈥檚 what they find.
You lay great store on imagination and on it linking with consciousness. Why?
I think that our contact with consciousness is mainly through our imagination. We spend more conscious time contemplating than acting. Contemplation is really a purposeful activity that involves imagination. If we only had perception of the world, we wouldn鈥檛 be so impressed with our consciousness. We can shut our eyes and imagine things, but even more than that we can read a book and live through the book. That is an amazing property and it鈥檚 probably unique to human beings. How perception and imagination are linked is the point of this work on depiction and inner states.
Are you the first to make a machine that can imagine?
I guess we鈥檙e the first. We published our first paper on it about a year ago.
How important is this in advancing progress towards a conscious machine?
Very important. In some sense, imagination is at the core of conscious experiences. It is far more intriguing than perception or images that have been remembered, because you鈥檙e creating something new.
The images your simulations produce are always from the machine鈥檚 point of view. Why?
The self-centred point of view is what we all have that is largely responsible for giving us the power to say 鈥淚 imagine鈥� or 鈥淚 see鈥�. A robot with a sense of self would have to start with this self-centred point of view.
The word consciousness conjures up not just imagination but intention, the ability to plan, to understand cause and effect. Have you come close to creating these?
I鈥檝e not come close really. But the point is that you start going in that direction and you look to see what it is that might block you. And we鈥檝e gone quite a long way. Even pretty hardened philosophers will accept that it is possible to see those capacities happening in a neural machine. Those are the easy bits about consciousness-to understand and simulate the neuroanatomy that gives us planning, a sense of time, of sequence and so on.
But this doesn鈥檛 get to 鈥渢he hard problem鈥� of consciousness-how do you know that NRM鈥檚 red is the same as mine?
I know that it isn鈥檛. It is very likely that if I look in your brain for the neurons that light up when you see a red ball, and then I look in my brain for those neurons, they might well be slightly different. The chances of you being able to predict from my neurons exactly what your neurons are doing isn鈥檛 possible. But we both agree it鈥檚 a red ball. What鈥檚 common to our two experiences is the ball. The realisation that the real world is the unifying standard between two people鈥檚 differing depictions suggests that the hard problem is a non-question. The easy problem is enormously hard and by the time we鈥檝e solved that we鈥檒l have got to grips with most questions about consciousness. And I think that鈥檚 possible.
So even if a computer became conscious it would experience the world differently from the way we experience the world?
Absolutely. It would be totally useless if it pretended to be a human being.
Apropos being a human being, Rodney Brooks at the Massachusetts Institute of Technology believes that you can鈥檛 have intelligent behaviours unless you interact with your environment . . .
What鈥檚 missing in his systems is the stuff that I do, the contemplative nature of systems. His systems are a bit like insects that respond to their environment directly. What鈥檚 missing in my model is the interaction with a real world. But we鈥檝e got around that by working in a very sophisticated virtual domain. We鈥檙e trying to get a project together with Caltech where they鈥檙e very good at building the sensory part of robots. We鈥檒l try and shift some of the contemplative stuff into robots.
Another approach is the work of Hugo de Garis at Star Labs in Belgium. You use artificial evolution to simulate small areas of the brain, he鈥檚 using it to create a whole brain. Is this feasible?
Hugo de Garis is very good at making large sheets of neurons. But he says all you need to make a brain is emergence and evolution. He鈥檚 trying to do everything-the lobes and the primary visual cortex and the eye. But what I don鈥檛 do is to try to evolve our systems from the primordial slime, which is what would be necessary if his system where to self-organise into a brain. It would probably take 50 or 60 million years. I wouldn鈥檛 go as far as saying it can鈥檛 be done. It could be done-perhaps.
If you did create a conscious machine would it need machine rights?
I believe not. What it would need is the engineering legislation which you have with any complex machine-curtailing what people might do with them. That鈥檚 the same as the control system of a plane, you know it has to pass certain stringent tests in order to make sure it鈥檚 not going to run amok. And that may be a bit difficult because these machines do arbitrary things, so there may be a problem with guaranteeing that these machines are safe. This brings us to Blade Runner and Do Androids Dream of Electric Sheep? territory. And, yes. Androids do dream of electric sheep.