I recently wrote about IBM’s Watson — a machine capable of competing against humans at Jeopardy!. The machine represents a pretty phenomenal leap in artificial intelligence, as it parses key bits of information out of natural language queries in real time. But here come the detractors, the “humanists” who are either too scared or too closed-minded to aknowledge the magnitude of this accomplishment.
In a piece called “What’s So Great About IBM’s Jeopardy!-Playing Machine?”, Niraj Chokshi argues that the machine isn’t actually “thinking” at all:
Rather than develop a machine that can decipher semantics — what do these words mean and how do they relate? — IBM took a shortcut of sorts and developed a high-speed computer that “thinks” in probabilities….
That’s quite a feat, but it’s not emulating human thought. It has side-stepped the problem altogether, relying on massive computing power and storage, as well as probabilistic number-crunching to approximate how we parse language.
And so we get to my favorite part of artificial intelligence – the philosophy. Too often, this element gets ignored or brushed aside but it really is fascinating. As we create devices that can “learn” (defined loosely as developing a consistent response to a stimulus) and “think” (combining stimuli – known or unknown – into new responses), we must investigate parallels with our own minds. Just as we learn about our own physiology from “lesser” organisms, so can we infer our mental processes from relatively simple models of learning.
But before we can address philosophy, let’s talk physics. Let’s think about Watson from the persepective of our own brains. Despite Chokshi’s claims, IBM isn’t “cheating” by using massive computing power. On the contrary, they’re playing catch up.
Estimates of the human’s brain capacity range from 1-1000 terabytes. It’s a wide range, to be sure, but for reference consider that the entire Library of Congress is estimated to contain only 20 TB of text. One-terabyte computer drives were only introduced three years ago; a cutting-edge supercomputer like Watson’s Blue Gene hardware might employ a database ranging from 600-1000 TB, on par with a human brain. A Blue Gene implementation scheduled for next year will use 1,600 TB.
But hardware alone does not a brain make! We still need software to access and interpret the data, as well as compression/access algorithms to store it efficiently. In this regard, modern technology doesn’t even begin to approach the human brain. Our compression algorithms are without peer – the brain stores only what it deems important and disregards extraneous details. Witness too how difficult it is to recall a perfect memory. It is well known that the brain uses cues to “fill in” what it doesn’t retain explicitly. So we may have the same storage capacity as Watson, but our use of that space is much more efficient, making it possible for humans to store far more data than the raw numbers suggest.
For further evidence, consider that Watson occupies thousands of square feet – your brain, needless to say, fits comfortably within your own head.
And how about speed (“massive computing power”)? Ray Kurzweil estimates that the human brain operates at about 20 petaflops (or 20 quadrillion operations per second). Blue Gene was designed to operate at petaflop speeds but currently sustains 0.5 petaflops. The latest computer to be designed has a theoretical speed of 20 petaflops, but will take years to reach that mark – and will not go into production until next year. So the human brain is operating at speeds an order of magnitude faster than IBM’s supercomputer, even before its superior access algorithms and storage capacity are utilized.
And as final evidence for the brain’s superiority, consider this metric of efficiency: your brain operates on about 20 watts. The next (and most power-efficient) Blue Gene supercomputer will draw 6 megawatts. That’s 6 million watts, or 300,000 times as much power as your brain.
Kurzweil has a nice graph that sums up the path of computers to human brain-like power:
Note that “Human Brain Functional Simulation” implies that the inner workings of the brain can be run at full speed – it does not mean any algorithms or intelligence will be present (presumably that must wait until 2025).
It’s silly to claim that Watson out-classes humans in speed, storage, or “smartness.” It is actually at a disadvantage because it lacks algorithms sophisticated enough to decide, at acquisition time, what information is important enough to retain and what can be ignored. On top of all those physical advantages, humans have some sense of “meta-learning” which we don’t yet know how to implement in hardware.
So we can turn to the second part of Chokshi’s claim – “probabilistic number-crunching [that] approximates how we parse language.” I wonder how he thinks we parse language, if not probabilistic number crunching? Like it or not, our brains are networks which pass and process information – they’re doing “math” in a very computer science-esque manner of speaking.
How does a computer know when it is right? It judges its response to be the most likely answer. Do humans act any differently? That’s “probabilistic”. How is that judgement performed? The answer is compared to other possible answers – perhaps faster than we are aware, but obviously the comparison takes place. As humans, we are fortunate to have replaced neurons firing with “sensations”; we don’t perceive that “x,y and z neurons fired – therefore “red” is the right answer.” Instead, we are overwhelmed by a sense of “red” and answer in kind. Computers don’t get the benefit of those “sensations.” They really do say “x,y and z factor receptors fired – therefore the answer is “red”". Masked or not, that is a “number-crunching” operation. (“Number-crunching” is rapidly becoming the crutch of people who don’t want to bother with understanding an underlying process. The term’s ubiquity is indicative of the prevalence of math in our daily lives.)
The actual mechanics of artificial intelligence are a bit beyond the scope of this article – but it’s safe to say they mirror cognitive processes closely, and a “thinking computer” isn’t too far off from a “thinking human” in terms of process and execution.
We could sum up this entire post by simply writing: “Your brain is an extremely powerful, extremely efficient supercomputer coupled with an extremely optimized unsupervised learning program.” But for some reason, that point of view is resisted by “humanists.” Is it fear of Skynet? Fear of playing God?
In order to truly understand these algorithms, we have to forget everything we think we know about computers as mindless data processors. These aren’t calculators, where an input leads strictly to a deterministic output. These are actual thinking machines which, given a set of input stimuli, weight and evaluate a set of likely outcomes — in some cases hundreds of times over — to produce informed guesses which, depending on the algorithm, might not even be the same each time it sees the data! Data processors (or calculators) require instructions at every step; thinking machines figure out the rules on their own. Make no mistake — computers are best at floating point operations (read: number-crunching); but artificial intelligence relies on moving away from simple deterministic outcomes and leveraging those fpo’s in a wildly different context.
In my mind, Watson is an incredible accomplishment. Moreover, it’s just a child – it’s intelligence is measurably a fraction of our own. In the coming years, these brains will become commonplace and we won’t fear them as competition – but rather as assistants. In the meantime, let’s not forget that the technology is in its infancy and WILL make mistakes – often ones which seem bizarre to our practiced brains.
Chokshi’s final argument takes a specific example which the computer got wrong and actually tries to make the argument that on that basis, the entire program is flawed. It’s a flawed tactic, to be sure, but since he claims that it is evidence of the computer’s failure, it seems reasonable to assume that he believes the reverse would be true of humans. Are we really so confident to think that just because information is “in our database”, we won’t give a wrong answer? If that were true, then humans with access to Google would never get a Jeopardy question wrong – as they would always, one way or another, be able to locate the “correct” information. Intelligence is not merely information storage and retrieval; it is the process of data interpretation.
In a recent dinner conversation about the philosophy of artificial intelligence, a friend pointed out that one of the theories I was discussing was almost a direct retelling of Plato’s theory of Forms. The idea was surprising, but reinforces the idea that artificial intelligence researches are not inventing a new discipline as much as they are trying to reinterpret a set of biological algorithms with modern hardware. I’m very excited to see what’s next.
