You can see both the chapters for this time as attempting to break down the idea that mental processes must be understood in Fodor's way, as involving rule-governed operations on mental representations that are stored in specific locations in the brain. Chapter 5 gives some reasons for thinking humans don't work this way; chapter 6 gives some reasons for thinking that robots don't have to be designed this way.
Fodor's picture, to review, involves chopping mental activity up into at least the following categories:
perception --> storage of sentences in the belief box
reasoning: rule-governed operations on the sentences in the belief and desire boxes; placing sentences in the intention box
sentences in the intention box --> physical movements
Chapter 5: Perception, Action, and the Brain
1. Marr's Three Levels
This is a nice, neat decomposition. Works great for digital computers. Does it work for the human brain? The assumption that it would has led some to downplay the importance of knowing about level 3.
Good thing about Marr's levels: Knowing about level 3 certainly isn't enough to understand how cognition works. For understanding, we need a more abstract understanding of what's going on.
Bad thing: it's not clear that things divide up this neatly in human reasoning (or need to in robot reasoning).
2. Biological Constraint and Liberation
Constraint: natural selection must build cognitive systems on top of whatever's already there. (Compare: lungs from swim bladders, panda's "thumb" from bone spur, human back from four-legged mammal back . . .)
Liberation: biological solutions do not neat to be neat and understandable, as long as they work.
Example 1: if you were engineering a hand-control mechanism, you might build whole-hand movements out of individual finger-movements. But in monkeys, individual finger movements involve more neural activity than whole-hand movements. Why? The whole-hand movements are more important, so there should be relatively easy triggers for those; then moving a single finger may require triggering the whole-hand movement and subtracting some of its components. (Or something like that!)
Example 2: The Fodor-style picture involves sharp distinctions between perception, cognition, and action. In practice, they seem to be messed up together. Adaptation to inverting or shifting lenses (p. 88).
3. Perception and Action: Messy Connections
1. No detailed inner models of full 3D scene. Brooks: "the world is its own best model." Just-in-time data collection. (Internet examples of huge things you don't notice. Car through intersection; dancing guy in monkey suit; color-changing card trick.)
2. Low-level perception may directly call actions that improve the perception. ("foveating" an object involves physical changes in the eye.)
3. Computation may itself involve real-world action. Calculating depth perception: can simplify the computation by moving head. (Head-bobbing and head tilting in animals.)
4. Representations of the world may be recipes for action rather than "passive data structures."
Chapter 6: Robots and Artificial Life
Cricket phonotaxis. How do crickets recognize the call of their own species, and figure out where it's coming from? The answer may not involve any internal representations at all. Automatically turns and moves toward sound of own species.
Last update:
March 31, 2008. |