Chapter 3: Patterns, Contents, and Causes
This chapter discusses three approaches to commonsense or "folk" psychology. Folk psychology tries to explain and predict human behavior by talking about "propositional attitudes" such as beliefs, desires, and intentions.
1. Fodor: the Representational Theory of the Mind
Fodor's view fits neatly with the classical approach to AI, and the "physical symbol system" hypothesis discussed in chapter 2. On this view, propositional attitudes like belief and desire are "computational relations to internal representations" (43), and mental processes such as reasoning are "causal processes that involve transitions between internal representations."
Here's a crude way to think about this. Suppose that there is a "language of thought." (This could be one's natural language, or it could be something more basic.) There might be one storage location (the "belief box") where we store sentences that represent our beliefs about the world, another box (the "desire box") where we store sentences that represent the ways we want the world to be, and a third box, the "intention box," where we store sentences that represent ways we intend to make the world.
Then theoretical reasoning would be a way of using the sentences in the belief box to derive further sentences, using (in part) deductive logic.
Similarly, practical reasoning would involve forming intentions based on the contents of both the belief and desire boxes. (For instance, suppose the desire box includes the sentence "I eat a hamburger," and the belief box includes the sentences "If I go to MacDonald's, I can buy a hamburger," and "If I buy a hamburger, I can eat a hamburger." I can use theoretical reasoning to derive the sentence "If I go to MacDonald's, I can eat a hamburger," and then I might use this sentence together with the desire to eat a hamburger to put the sentence "I go to MacDonald's" in the intention box.
(Then there will be processes that lead from perceptual inputs to sentences getting added to the belief box, and other processes that lead from sentences being included in the intention box to actual actions.)
Very very crudely speaking, this is the way some AI programs work, and Fodor's idea is that we work this way too.
2. Churchland: Eliminative Materialism
Churchland thinks that the whole vocabulary of beliefs and desires is so hopelessly misguided that we should throw it out and start over. In particular, he thinks that there's nothing in the brain that relates very closely to our pretheoretical notions of belief, desire, etc., and therefore we should reconstruct our vocabulary for the mind so that it corresponds more closely to what's actually happening in the brain.
For C, this would be part of a very ambitious project of thinking about the world in ways that are scientifically accurate instead of in commonsense ways that are seriously misleading. He thinks we can do this with regard to the external world: we can learn to see the earth as revolving around the sun, we can learn to think about mean molecular kinetic energy instead of "heat," and so on. His eliminative materialism about the mind is in part a recommendation that we do the same thing with regard to the internal world.
3. Dennett: the Intentional Stance (Instrumentalism)
Dennett is agnostic about whether Fodor or Churchland is correct about whether the brain actually contains mental representations that are manipulated in accordance with rules. But unlike both Fodor and Churchland, Dennett thinks that belief-desire psychology is useful and important regardless of whether or not we can find mental representations in the brain, so to speak.
Dennett is often called an "instrumentalist," because he regards beliefs, desires, etc. as useful fictions. (This is the way the early 20th century instrumentalists regarded things like electrons and protons.) For D, to say that people have beliefs and desires is basically just to say that it is useful to treat us as though we do. (It's useful to take up the "intentional stance" with regard to people.)
Dennett distinguishes between three ways we can consider things:
From the physical stance, we think about things in terms of the physical laws that govern their behavior, and we use those to make predictions. From the design stance, we think about things in terms of how they are designed to operate, and we predict that they will do what they are designed to do (either by an intelligent designer, or by natural selection) in the circumstances they are actually in. (I don't need to know how my thermostat works to figure out what it will do if it's set at 78 and the termperature drops to 76.) From the intentional stance, we think about things as having beliefs and desires, and as doing what is rational under the circumstances.
What will Fred do if I ask him what 125 + 17 is? Normally I just expect him to say 142, since that's the rational answer. If he says 132, I may need to drop down to the design level and consider what "program" or "algorithm" he's using in order to figure out where things went wrong. (Maybe he's using a defective addition algorithm that leaves out the 'carry the one' step.) If he just stares blankly at me and makes weird noises, I may need to drop all the way down to the physical level and look for physical causes for his odd behavior.
Dennett suggests that AI programs can be intentional agents, because the most useful way to predict what they'll do can be to assume that they have beliefs and desires. (What will the chess playing program do? We may get better results by noticing that it "wants" to get its queen out early than by actually trying to trace through how the program works.)
Chapter 4: Connectionism
Connectionism (neural nets, parallel distributed processing) is a very different approach to AI.
The idea is to model AI programs more closely on how the brain works at the level of individual neurons and their connections. So we'll have layers of "neurons" with connections between the layers. We'll have a layer of input neurons, one or more layers of "hidden" neurons, and a layer of output neurons.
Connections can be either excitatory or inhibitory (that is, firing one neuron can make it either more or less likely that the neurons it's connected to will fire). The connections will have weights representing how strong the excitatory or inhibitory influence is.
Then we "train in" a network: assign weights randomly at first, and use a back-propagation algorithm to modify the weights whenever a wrong answer is given.
Interesting fact: in classical AI, problems that humans find difficult (e.g. multiplying large numbers quickly) are easy, and problems that humans find easy (recognizing faces, e.g.) are difficult. In this regard, connectionist AI seems to work more like people: neural net programs are good at pattern recognition, but less good at some of the things traditional AI found relatively easy.
Last update:
March 26, 2008. |