5.9 Prediction Machines

• William James: Try to feel as if you were crooking your finger, whilst keeping it straight. In a minute it will fairly tingle with the imaginary change of position; yet it will not sensibly move, because ‘it is not really moving’ is also a part of what you have in mind. Drop this idea, think of the movement purely and simply, with all brakes off; and, presto! It takes place with no effort at all. [The Principles of Psychology, 1890, p527.]

Everyone can think about things, without performing actions—as when Carol imagined moving those blocks. But how did she manage to do that? You, yourself could now close your eyes, lean back in your chair, and indulge in your own dreams and fantasies, reflect upon your motives and goals, or try to predict what will happen next.

Now, here is how we could make a machine that does that same sort of thing, by predicting the outcomes of various actions. Let’s assume that it has some rules like these.

Then we’ll give our machine—let’s call it Seer—a way to replace what it currently sees by the prediction described by this rule. Then when Seer is in situation A, and then considers doing action X, this will cause Seer then to 'imagine' that it is now in a situation like B.

I included that pair of “Suppressor Bands” for two separate reasons. First, when Seer imagines that future condition B, we do not want this to be quickly replaced by a description of the actual, present condition A. Second, we do not yet want Seer to perform action X, because it might want to consider some other options before it makes a final decision. So, Seer can use those suppressor bands to detach itself from the outside world—and this enables it to “stop and think” before it decides which action to take.[16]

By repeating this kind of operation, Seer could use such prediction-chains to simulate what happens in ‘virtual worlds.’ Of course, for Seer to be able to make such predictions it must be able to use the kinds of search we described in §5-3 to simulate (and then compare) the effects of difference courses of action before deciding which one to adopt. This will need additional memory, as well as other kinds of machinery. Still, anyone who has played a modern computer game can see how advanced has become the art of building virtual worlds inside machines.

I expect that in the next few years, we’ll discover structures like those in this diagram in various parts of human brains. How did our brains evolve these abilities? The species of primates that preceded us must have had some structures like these, which they could think several steps ahead. But then, a few million years ago, that system appears to have rapidly grown, as the frontal lobes of our brains developed their present great size and complexity—and this must have been a crucial step toward the growth of our human intelligence.

Summary This chapter described some structures and processes that might do some of the things that people do. We outlined a sequence of levels at which we can use increasingly ways to think

However, we have suggested rather few details about what happens at each of those levels. Later I will suggest that our systems mainly work, at each of those various cognitive levels, by constantly reacting to the particular kind of troubles they meet—by switching to more appropriate Ways to Think. We’ll represent this Model of Mind by using this simple diagram:

The “Critic-Selector” Model of Mind

In the rest of this book we will frequently switch between these two different views of the mind—because each one gives better answers to different kinds of questions about ourselves. Model Six makes better distinctions between various levels of mental behaviors, whereas the Critic-Selector view suggests better ideas about how to deal with difficult problems. Chapter §7 will combine both views, because we frequently use different Selectors and Critics at each of those various cognitive levels.

However, no matter how such a system is built, it will never seem very resourceful until it knows a great deal about the world it is in. In particular, it must be able to foresee some of the outcomes of possible actions, and it won’t be able to do this unless until it possesses the right kinds of knowledge. For human beings, that’s what we mean by “commonsense” knowledge and reasoning. And although, in everyday that phrase means, ‘the things that most people find obvious,’ the following chapter will demonstrate that this subject is surprisingly complex.

[1] More details about construction planning were developed by Scott Fahlman in his 1973 paper at ftp://publications.ai.mit.edu/ai-publications/pdf/AITR-283.pdf

[2] In Principles of Psychology, p359

[3] According to Tinbergen, when an animal can’t make a decision, this often results in dropping both alternatives and doing something that seem to be quite irrelevant. However, these "displacement activities" seem to be fixed, so they do not suggest that those animals have thoughtful ways to deal with such conflicts.

[4] In The Natural History Of Religion, 1757. http://www.soci.niu.edu/~phildept/Dye/NaturalHistory.html

[5]Some early steps in that project are described in ftp://publications.ai.mit.edu/ai-publications/pdf/AIM-200.pdf.

[6] See http://web.media.mit.edu/~minsky/papers/PR1971.html

[7] In fact, that darker horizontal streak is not the lower edge, but is part of the surface next to that edge, slightly shadowed because that edge is worn-down.

[8] V.S. Ramachandran, Science, v305 no.5685, 6 August 2004.

[9] in www.richardgregory.org/papers/brainy_mind/brainy-mind.htm. See also, www.physiol.m.u-tokyo.ac.jp/resear/resear.html

[10] This program was based on ideas of Yoshiaki Shirai (and Manuel Blum). See ftp://publications.ai.mit.edu/ai-publications/pdf/AIM-263.pdf. However, I should add that Builder had almost no competence for any but neat geometrical scenes—and, so far as I know, there still are no ‘general-purpose vision machines” that can, for example, look around a room and recognize everyday objects therein. I suspect that this is mainly because they lack enough knowledge about real-world objects; we’ll discuss this more in Chapter 6.

[11] See papers by Adolfo Guzman and David Waltz at ftp://publications.ai.mit.edu/ai-publications/pdf/AIM-139.pdf and

ftp://publications.ai.mit.edu/ai-publications/pdf/AITR-271.pdf

[12] See Zenon Pylyshyn, http://ruccs.rutgers.edu/faculty/ZPbbs98.html. [Broken Link] The octagon example is from Kanizsa, G. (1985). Seeing and Thinking. Acta Psychologica, 59, 23-33.

[13] In this kind of diagram, each object is represented by a network that describes relationships between its parts. Then each part, in turn, is further described in terms of relationships between its parts, etc.,—until those sub-descriptions descend to a level at which each one because a simple list of properties, such as an object’s color, size, and shape. For more details, see §§Frames, Quillian’s thesis in Semantic Information Processing, and Patrick Winston’s book, The Psychology of Computer Vision.

[14] Some persons claim to imagine scenes as though looking at a photograph, whereas other persons report no such vivid experiences. However, some studies appear to show that both are equally good at recalling details of remembered scenes.

[15] See , for example, http://www.usd.edu/psyc301/Rensink.htm and http://nivea.psycho.univ-paris5.fr/Mudsplash/Nature_Supp_Inf/Movies/Movie_List.html.

[16] This prediction scheme appears in section §6-7 of my 1953 PhD thesis, "Neural-Analog Networks and the Brain-Model Problem, Mathematics Dept., Princeton University, Dec. 1953. At that time, I had heard that there were ‘suppressor bands’ like the one in my diagram, at the margins of some cortical areas. These seem to have vanished from more recent texts; perhaps some brain researchers could find them again.