Computational Mind, Conclusion.

Ten years after ENIAC, the discipline of artificial intelligence was born. At the time, computer scientists thought we could create a computational intelligence. Computers were doing things that no one had imagined before. John von Neumann (who I am convinced was an alien) had introduced the theory of self-sustaining automata a few years before. Alan Turing had developed the Turing Test for cognitive systems. Neural network theory had been around since 1943. Everything was in place.

So what happened?

A chasm opened up between understanding the brain and understanding the mind. The more the AI scientists discovered, the wider and deeper it got. We could do a lot of the brain's cognitive functions, but that's just mechanics. We couldn't create a computational mind. In parallel, neuroscience was discovering the same thing, that brain is different than mind. Marvin Minsky, the godfather of artificial intelligence expressed doubts about whether the human brain was able to comprehend itself.

Computation means applying a set of rules to a problem to solve it. In principle, discovering the rules that apply to the human mind is no more difficult to do without a computer than with one. The computer relieves the tedium of trial and error and speeds the process, but, because the computer's rule system is built by us, it will not reach a more surprising conclusion. There are two major issues that will have to be addressed: learning and emotion.

We learn every single instant of every single day. It was colder than we expected. That song sounds different in another room. The curry was spicier than we wanted it to be. It wasn't spicy enough. And, every little quanta of new information changes our cognition. That color, smell, and texture of curry will taste a certain way. Until it doesn't. We can simulate learning, but we can't create a machine that actually learns. Or, at least not in any meaningful way. One AI scientist put it that a machine intelligence needs to be able to learn from testimony. That is, it needs to be able to read a book and learn from it.

Emotion colors everything we experience and everything we do. We are not, nor will we ever be completely logical and rational beings. We're depressed, so the food doesn't taste as good, the colors aren't as vibrant, the smells are not as pungent. We're happy and everything is better. Emotion affects our learning as well. Our mood enhances or detracts from what we learn of our surroundings as well. If we are depressed or angry or any one of a litany of negative emotions, we are less likely to notice our environment, or we are more likely to focus on just one or two elements of it. An emotionless machine will never successfully interact with humans over the long term. It may far exceed our cognitive abilities, but it will never understand us or the world as we see it.

Furthermore, learning and emotions are not binary. This represents a serious problem for machines that are based on ones and zeroes, on or off, yes or no. There needs to be a third state. Yes, no, or maybe. Actually, there need to be a number of intermediate states between yes and no. Consider the well used paradoxical question, when did you stop beating your wife. There is no binary state. Any binary answer is wrong if you never beat your wife. You need the third state, I never beat my wife, or even a fourth state, why are you asking me that question. And, the latter illustrates a fundamental problem with a binary intelligence. Questioning the question. A machine that always wants to know why you asked the question would not be useful and would be very frustrating. A learning machine would need to know when to stop questioning the question.

It would be even more frustrating if the machine always gets stuck on a paradox. Ask me what would happen if I went back in time and killed my grandfather. The paradox is that I would have never existed to go back and kill him. We might ponder it for a minute or two and move on. The mind balks, so we just go around it. A rule-based, computational mind would get stuck if there was no escape mechanism. If we build the escape, then we risk the machine's ability to recognize the paradox in the first place. It would need to learn to recognize the paradox and go around it.

Most of you reading this have seen, or at least are aware of a movie called The Terminator. In it, a highly interconnected network called Skynet was built by the military. It used the idle time of thousands of computers to create a virtual CPU. It's an interesting idea, and not all that unlikely in principle. SETI (Search for Extraterestrial Intelligence) runs a program that allows Internet participants to donate their spare CPU time to a huge star mapping effort. There is also a similar protein folding effort hosted by Stanford University. When Skynet became self-aware, it evaluated humanity, declared us vermin, and went about the business of exterminating us. This is not unlikely either. It is a legitimate concern expressed by some of the artificial intelligence community. A computational mind could potentially be a superior being, and there is no guarantee that its intentions would be completely benevolent. There are ethical concerns as well. What are the ethical consequences of flipping the power switch on a machine intelligence? Mind and body are symbiotic. Killing a human being means killing its body as well as its mind. When the body is irreparably damaged, but the mind continues to function, we don't flip the switch. The ethical dilemma already exists. We certainly don't euthanize quadriplegics.

Fortunately for us, we don't have to face the nightmares of Skynet or the ethical dilemmas of flipping the switch on a living machine mind. Current AI science says that the best hope for a fully functional computational mind will be to simulate the full neuronal network of the brain. That is, neurons and synapses. That is an incredibly daunting task. Last year, the Japanese K computer was used to simulate a small neuronal network. It used over 82,000 high powered processors to simulate 1.73 billion nerve cells connected to 10.4 trillion synapses. That's about 1% of the average human brain. It took the computer 40 minutes to simulate 1 second of biological activity. The extrapolation of CPUs and time is not linear. To double the number of neurons and synapses, it would take much more than double the number of CPUs and time. Simulating one day of biological activity for 20% of a single human's brain capacity would consume all of the processing power on the planet for years.

The androids depicted in science fiction that look like regular humans are not likely any time soon. While the carriage of the robot could be largely humanoid, the brain housing would have to be the size of a small aircraft hangar. Our current day android would also need to haul a couple of tractor trailers around for the 10 megawatts of power it would consume (about 10,000 suburban homes) and the air conditioning to cool it. Moore's law says the number of transistors on an integrate circuit roughly doubles every two years. (It's not really a law, it's more of an observation.) It would take about a century to reduce the K to the size of a human skull and it would still only encompass 1% of a normal brain's topography.

But we can still dream, can't we? And don't dreams often precede facts in science?