"But though our thought seems to possess this unbounded liberty, we shall find, upon a nearer examination, that it is really confined within very narrow limits, and that all this creative power of the mind amounts to no more than the faculty of compounding, transposing, augmenting, or diminishing the materials afforded us by the senses and experience."
-- David Hume, An Enquiry Concerning Human Understanding, section 2 Jeff Hawkins, the inventor of the Palm Pilot, reminds me of eighteenth-century Scottish Enlightenment philosopher David Hume.
In an important new book, On Intelligence, written with assistance from Sandra Blakeslee, Hawkins takes the view that abstract thinking and sensory perception are aspects of the same mental process.
Hawkins believes that by studying the architecture of the brain we can learn better ways to build intelligent machines. Accordingly, his book tries to provide the reader with a state-of-the-art view of how the brain functions.
Hawkins focuses on the brain as constantly trying to establish patterns and using trial-and-error exploration to fit those patterns to real-world experience. In addition to being inherently interesting, this description intrigued me, because I have found trial-and-error learning to be central both to my career as an entrepreneur and to the way I characterize the economy as a whole.
Let me try to illustrate Hawkins' central point using as a metaphor the game Hangman, a word game in which you guess letters one at a time. Suppose that for a five-letter word you have guessed the letters B, T, and U, with this result: _ _ _ U T. That is, there is no B in the word, but it ends in UT. What letter would you guess next?
A good strategy at this point is to try to think of words that end in UT but do not contain B. Possibilities include "spout," "clout," "snout," and so on. This makes an O or an S a good letter to guess. On the other hand, it is hard to think of a word that would fit that has an A or an E. Even though those letters are common in general, they would be poor guesses in this instance.
In Hawkins' model, the words that we have in mind feed into our decisions about what to guess. The results from the guesses then give us more information to make our next guess.
The processes of guessing and getting new data work together. Hawkins argues that sensory perception uses this process of pattern-based guesswork. For example, as our retina obtains visual data, the cortex sends back a pattern, such as a friend's face, to help the guide the retina's next activity. The retina's signals are like the results from our hangman guess, and the friend's face is like the word that we hypothesize is the answer. Just as we may have the wrong word, we may have mistakenly identified our friend. As we get more data, we may decide that it is not our friend's face that we see.
Hawkins argues for the centrality of patterns in thought. The brain constantly forms patterns. It communicates internally through patterns. This facility with patterns is what Hume described as "compounding, transposing, augmenting or diminishing" our sensory data. In fact, Hawkins would argue that we experience our sensations as patterns. For example, what we see is necessarily a combination of the patterns predicted by our brain and the signals sent from our retina.
Hawkins suggests that the brain architecture that is needed to use pattern-matching to process sensory information is exactly the same as the architecture needed to think in terms of abstractions. For example, whether you read these words on a screen or hear them read to you, your mind uses its generic pattern for language recognition to process the information.
At one point, Hawkins says to imagine that you were asked to recall the opening lines of Lincoln's Gettysburg Address. Your brain would activate the pattern "Four score and seven years ago..." Then, if you were to recite your answer, your brain would activate the patterns that control your speech. On the other hand, if you were writing your answer with a pen and paper, your brain would activate the patterns that control your hand muscles. The "Four score" pattern can be used either way. Intelligent Machines?
What does this model of human intelligence imply for those who are attempting to develop intelligent machines? Hawkins writes (p. 209), "Here, then, is the recipe for building intelligent machines. Start with a set of senses to extract patterns from the world...a set of senses that differ from a human's...Next, attach to these senses a hierarchical memory system that works on the same principles as the cortex. We will then have to train the memory system much as we teach children. Over repetitive training sessions, our intelligent machine will build a model of its world as seen through its senses. There will be no need or opportunity for anyone to program in the rules of the world, databases, facts, or any of the high-level concepts that are the bane of artificial intelligence."
I agree with Hawkins that a machine that can learn in this way would be more intelligent than a machine that requires a database or a set of rules. However, I am not sure that I agree that the key to creating such a machine is to mimic the brain's pattern-recognition architecture. It seems to me that Hawkins' theory of intelligence does not warrant the conclusion that with a better wiring diagram and more memory we can create a machine that learns.
My computer has a much bigger word dictionary than the average eight-year-old. Yet it would be easier for me to teach an eight-year-old to play Hangman than to teach my computer to play the game. That is because the computer is too stupid to realize that it even has a dictionary. The dictionary is stranded inside a word processing program, with no way to get out. In fact, there are a number of "spell-check" dictionaries in my computer, all completely unaware of one another and totally oblivious to my interest in playing Hangman. Instead, if you want to teach my computer to play Hangman, you have to give it a separate program that either "learns" a dictionary by repeatedly playing Hangman or takes a dictionary from the programmer.
The Kling Test
In fact, I propose a test for artificial intelligence -- my version of the Turing Test, if you will. When a computer is smart enough to figure out without specific instruction that the spell-check dictionary in its word-processing program can be useful for playing Hangman, then the computer meets the Kling Test. An intelligent computer could figure out how to search its dictionary for five-letter words ending in UT without having to wait for someone to write a PERL script to do so.
Hawkins uses another example that illustrates this point -- teaching a robot to catch a ball. This turns out to be a really difficult exercise in hardware and software engineering. Yet a child can learn easily, because: (a) the child has experience in grasping things, intercepting things, and other similar movements; and (b) (Kling test) the child can figure out how to translate those experiences to the new problem of catching a ball.
I came away from reading On Intelligence with a relatively pessimistic view of the potential for machine intelligence. Hawkins' book makes Ray Kurzweil seem simplistic for implying that sheer hardware power is going to enable computers to achieve human intelligence. Indeed, I think Hawkins himself under-estimates the difficulty of the task of creating machine intelligence, given his model of the brain.
I am struck by two aspects of the human pattern-creation process: its dependence on sensory stimulation; and the importance of synthesis, of applying patterns to new situations -- once again, Hume's "faculty of compounding, transposing, augmenting or diminishing."
Compared to humans, a machine lives in a sensory-deprivation chamber. The machine is not receiving the volume and variety of sensory input that humans receive. Hence, it is not going to be able to form very interesting patterns.
Secondly, a machine does not face the variety of tasks and problems that a human faces. Even the machines that Hawkins describes, such as an intelligent automobile, would have a very narrow focus. Thus, their capacity to synthesize patterns would be severely limited. Such machines would be idiot-savants, not intelligent entities.
When we get beyond sensory and motor skills and proceed to the level of abstract reasoning, then it seems to me that the patterns might be analogies. We approach a new concept by treating it as a combination of analogies with existing concepts. For example, there are many board games that include moving pieces based on a roll of dice and following directions written on cards. If we have played Monopoly, then other games will be more quickly grasped by analogy.
Perhaps we form new concepts by creating new combinations of analogies. It could be that problem-solving and creativity involve trying new combinations of analogies and evaluating whether they bring us closer to a goal.
I think that if I were trying to build a learning machine, my strategy would not focus on hardware. I would try to build a machine that could develop and manipulate a database of analogies. I would probably try to come up with (or borrow from someone else in the field) an "algebra" for working with analogies that the computer can use in order to construct and test new patterns. That would be an old-fashioned approach, in Hawkins' view. However, after reading On Intelligence, I think that the key to getting any machine to learn is to give it a variety of both stimuli to absorb and tasks to perform. Moreover, it is important to have the machine synthesize its knowledge, rather than use a separate program for each task.
I learned a great deal from On Intelligence. If you found this essay interesting, then I strongly recommend reading Hawkins' book yourself. My personal take-away from On Intelligence is that humans have an advantage in that our database of analogies starts out with all of the patterns that come from our sensory and motor system. Moreover, we naturally synthesize our brain functions, rather than function with highly separate programs as in a computer.