Apr 30 2007
Neuroscience and computer technology both deal with the underlying science of information storage and processing. The brain, after all, is a complex, massively parallel, biological computer. At the cutting edge, these two disciplines are benefiting from a robust cross fertilization: neuroscientists are using computers to model portions of the brain and better understand how it works; computer scientists are using neural network models to create computer programs that learn. The ultimate expression of this overlap is to produce a computer that models a biological brain. A team of researchers from the IBM Almaden Research Lab and the University of Nevada have made significant progress toward this goal.
The team has used IBM’s BlueGene supercomputer to simulate a mammalian brain with about half the computational power of a mouse brain (8,000 neurons each with about 6,300 synapses, or connections), and operating at about 1/10 the speed. They claim that when they ran their virtual mouse brain they saw virtual neurons form spontaneously into groups and nerves firing with staggered coordinated patterns typical of a biological brain.
There are two basic approaches to computer simulation of a brain. The first is to design the physical computer after the anatomy and function of a brain. Our brains differ from current computer technology in several important ways. Brains are analog devices, while computers are digital. The analog nature of biological brains is due to the fact that neurons connect to each other with varying strength – as opposed to a digital system that uses absolute “on/off” binary signals, brains have analog signals that can vary along a continuum in terms of strength.
Second, our current computer technology is largely serial – computations are made one after another. We are just starting to use parallel processing to a significant degree, although supercomputers have used parallel processing longer than desktops. But even the most parallel supercomputer is nothing compared to the human brain, which is massively parallel. Each computation may be much slower in a human brain, but the brain can make billions of such calculations at the same time.
Third, biological brains have a physical mechanism for learning. Life is very good at growing and adapting. Silicon is not so good at this. Brains grow more and stronger connections to reinforce pathways. Computer chips cannot do this.
So for now we are far away from building a computer that replicates the key physical processes of a biological brain. It is not even clear at this point what direction future technology will take to accomplish this. We may even design biological computers that use living components in their designs.
Fortunately, there is a second way to simulate the properties of a biological brain – with software rather than hardware. This is the method used by the IBM Almaden Research Lab. Essentially, they run the simulation on a regular digital computer but the software is designed to simulate a virtual brain. This is inefficient, in that it takes a great deal of digital processing to reproduce the analog parallel processes of a biological brain, but it has the advantage of using existing digital computers. Also, silicon (and likely whatever succeeds silicon) is much faster than biological neurons, and eventually we will have fast and powerful enough digital computers to simulate the processing power of a human brain.
Obviously, we still have far to go. If Moore’s law holds up (and this seems likely) we should have computers with enough processing power in about 20 years. And of course computers will only become more powerful from there. It is harder to predict how long it will take to design the software to create the virtual brain, but what this latest research shows is that this is an area of active research that is making significant progress. It is plausible that the software research will keep pace with hardware developments, and in a couple of decades or so we will have artificially intelligent computers operating with the complexity and power of a human brain.
The research team promises that their next generation software will be more complex and adaptive, more closely resembling a mouse brain than their current version. I have no doubt this is true. There are really no theoretical obstacles to developing virtual brains.
One philosophical aspect of this technology intrigues and confuses me, however. I have no problem imagining artificial intelligence based upon silicon that is as self-aware and conscious (whatever that really is) as a human, if the AI computer is physically designed to replicate the properties and processing of a biological brain. But what about a virtual brain, like that being developed. Can a virtual brain be self-aware, or is it just a very good mimic of AI? I know this is a generic conundrum of all AI, yet it doesn’t bother me with computers that physically create AI and it does with those can virtually create AI. Perhaps it shouldn’t.
Here’s my problem. If we take a digital supercomputer that is not designed for AI – it’s just an ordinary digital computer, but very powerful – I think we would agree that it is not self-aware. But then we run virtual brain software on the computer, does the computer become self-aware while the software is running? Or is the software self-aware and not the hardware? And what if I close the virtual AI program – does the AI die? And if I restart the program later is their continuity to the previous virtual being, or did I just create a new AI?
This is not a new debate – I think it’s just the same soft AI vs hard AI debate that has been around for decades. Is information the only thing that is necessary to have self-aware intelligence, or do you need hardware that is designed to carry out the processes of self-awareness? I tend to fall into the latter camp, but I confess this is more an intuition than solid logic.
Leave a Reply
You must be logged in to post a comment.