Slower Artificial Intelligence
Most involved in artificial intelligence research and practice would like faster computers and software. The basic premise remains to be that the brain is a very fast and efficient computing machine and to replicate intelligence will require significant improvement in speed and size of computing. This makes sense. However, such a process is a linear extension of traditional computing ideas in an effort to explain complexity, that is not well understood.
It is conceivable that the brain is not a fast computing machine at all. However, it demonstrates the ability to assimilate complex information quickly to produce outcomes that are extremely impressive. There are two assumptions in the previous statement – first, the brain is assimilating complex information and second it is producing outcomes that are impressive. The first assumption is an observation that the brain has many channels through which it appears to be collecting information but it is not exactly clear if it is using all that information. An alternative hypothesis may be that brain is using only a small percentage of the available information. If the discarding of information is systematic, it will require high computing power but it is also possible that the brain is using simple rules of thump and indiscriminately discarding information because it simply does not have the computing power to process it.
The second assumption that the outcomes the brain produces are impressive is debatable– the human brain perceives its own process as impressive but that is not an absolute. For example, if the brain is selecting from available behavior and outcomes templates (that are limited) based on the inputs, the fact that the templates are complex does not mean that it requires computing intensity. Artificial Intelligence focuses on creating the templates from scratch based on a large amount of input information but the human brain may not do that at all. The operating system, at birth, is loaded with some basic templates and over time the brain makes some changes to these templates. But the outcome itself is really a selection problem (i.e. which template to produce) rather than a design problem (how to design the template).
Many have been fascinated by the workings of the human brain and many have been toiling for the past few decades to try to replicate it using traditional computing – by making the machines faster and their memory larger. It may be worthwhile to step back and challenge the basic assumptions underlying this effort. It may be possible to reproduce the outcomes of the brain by simple processes that require low computing intensity and memory. If so, such intelligence may not be worth replicating.
We think of the human brain’s processes as impressive not because they are particularly difficult, but because we don’t understand them. Bird flight was very impressive before planes were around, but it’s sufficiently simple to construct a paper airplane, or understand the principles of flight.
The fundamental disconnect between artificial intelligence in a computer and intelligence in a human is more probably a result of differences in logical processes and data representation, not computational power.
For example, if I tell you that I’m thinking of a bird, then somewhere in your brain you might conclude that the thing I’m thinking of flies. However, if I later tell you that I’m thinking of a penguin, you will reevaluate your conclusion, but you were not wrong in the first place, because the archetypal bird does actually fly. On the other hand, any computer program that is claimed to be artificially intelligent (for instance, IBM’s Watson) would be quite incorrect to come to the same conclusion, because that wouldn’t be useful to us. Watson is not supposed to make mistakes because it has all the data!
This brings up another point, that the data we’re computing in our brains is vastly different from the data in a computer. For instance, my profile icon is represented as a string of numbers that would easily fill a printed page of paper. Every pixel is precisely spoken for, and there is no ambiguity as to what the picture looks like. The same picture recollected by my brain is flawed. I don’t recall exactly what shade of grayish-brown the tree is, and I don’t actually care. Would anybody want a computer to do the same?
So while we can speculate that a brain’s processes might be simple, we still don’t remotely understand them, and artificial intelligence research as it is today focuses much more strongly on the kinds of problems with clear-cut answers. We are still leagues away from anything resembling a human brain in software.
I am grateful for your comments. Every time I write a blog entry, I have the Pink Floyd song in my brain, “Is there anybody out there?.. It appears that there is somebody out there.
I agree with most of what you say – perhaps all of it – but I will restate in the following way.
(a) Humans are distinctly handicapped analyzing their own brains although most humans appear to be fascinated by it and for good reasons.
(b) I agree that the notion of “artificial intelligence” is flawed – as it appears to represent vast amounts of data storage and retrieval (e.g. Watson).
(c) I agree that we do not understand the brain – but that does not mean that brain is housing any major secrets – it may be just that we are not capable of understanding it. In other words, the fact that we do not understand the brain does not mean that the brain is complex.
So, I think, in general I agree with you. I think the human brain is a relatively simple construct.
I’m actually not interested in whether the brain is complex or not, and if compelled to pick I’d probably default to the traditional view that it is very complex. My point was just to discriminate that most artificial intelligence is hardly related to imitating the brains processes. At least, it used to be and people gave up on it for the most part in favor of more tractable results.
The methodologies involved in writing AI software have more to do with the mathematical representations of the data. At best, AI mimics some processes we observe in nature, like ant colony optimization, but even research in so-called “neural networks” has not progressed since the 80′s, in favor of statistical methods to give approximately good solutions.