Class Notes: Chapter 2

Thinking Machines

Comparing ELIZA and SHRDLU: If one is a better model of human behavior than the other, on what grounds does one make that judgment?
Consider the common criteria for evaluating theories: Testability, Accuracy, Generalizability, Parsimony

Chapter 2: Section 2.1

In what ways does SHRDLU go beyond a chatterbot such as ELIZA?

What does SHRDLU “know”?
What does ELIZA “know”?
What skills does each possess?
Which is more human? In what way?

Give three reasons why SHRDLU was important in the birth of cognitive science.

SHRDLU uses natural language as the mechanism for input and output
SHRDLU models abstract grammatical rules
SHRDLU uses task analysis to identify distinct components, and to suggest algorithms for accomplishing those tasks

Is it worthwhile for cognitive scientists to try to build machines that mimic cognition?

Consider: “To understand something is to be able to recreate it”. Is this ability really a necessary or a sufficient prerequisite? How else can we test our understanding?

Is it a concern that SHRDLU only deals with a micro-world and a restricted language?

What is it that is missing in SHRDLU?
Conversations are social activities, with many implicit assumptions, and depends on the broader context of the conversation
This is not necessarily a weakness in SHRDLU. We do not seek a “Theory of Everything”

Schank & Riesbeck: SAM - A Program that Uses Scripts

John went to a restaurant. He ordered a hot dog. The waiter said they didn’t have any. He asked for a hamburger. When the hamburger came, it was burnt. He left the restaurant.

Q. Did John sit down in the restaurant?
A. Probably.
Q. Did John order a hot dog.
A. Yes.
Q. Did John eat a hot dog?
A. No, the waiter told John the management was unable to give it to him
Q. What did the waiter serve John?
A. The waiter served John a hamburger.
Q. Why didn’t John eat the hamburger.
A. Because the hamburger was overdone.
Q. Did John pay the check?
A. No, John was angry because the hamburger was overdone and so he left the restaurant

How does SAM differ from SHRDLU? What can SAM do that SHRDLU cannot?

SAM has a script-based knowledge of the real world
Schank was in large part responsible for introducing the theory of scripts
The scripts are still limited - SAM knows about restaurants, but not about many other things

Chapter 2: Section 2.2

A concern with information processing inevitably brings up the issue of representation

The associationist tradition:

Began with Aristotle, implicit in behaviorism
Cognitivists assumed that information was represented by associations among ideas

What is the question at issue in the mental imagery debate?

Associations are inevitably digital in form. Is there a need for something else?

How do digital and analog representations differ from each other?

Is analog closer to reality?
Consider the watch
What are the advantages to a digital watch?
What are the advantages to an analog watch?

Evidence for Analog Representations (Brooks, 1968)

Form a memory image of a capital F
Trace around the image, starting at the bottom left corner and working clockwise
Indicate for each corner whether it is on a top edge of the figure
Performance is impaired when responses are made visually (i.e. by pointing to the word ‘Yes’), rather than by saying ‘yes’
What does this imply about the representation?

What was the contribution of Kosslyn’s experiments and Shepard and Metlzer’s experiments to the mental imagery debate?

Previously, discussions of mental images seemed to be too subjective
These studies connect an observable behavior (response time) to hypotheses about mental processes
The assumption: Operations on mental images follow the same principles as operations on real objects

Is introspection a valid method in psychology? With respect to Shepard and Metzler’s experimental paradigm, is it significant that it seems to participants as if they are rotating one image to compare it with the other?

There are many reasons not to trust our intuitive sense of how the mind works
Evolution is a cheap skate – it does not purchase unnecessary skills
Musacchio (2005): Consciousness works on a “need to know basis”
If we do not need to know how the mind accomplishes something, it does not bother to inform us

Further Evidence for Mental Images

Imagine (a) a large elephant next to a small mouse, or (b) a small elephant next to a large mouse
“Does a mouse have whiskers?” What would you predict?

The critique: Do we need two forms of representation? The principle of parsimony: Do not propose two separate systems if one will do.
Numerous critics of the imagery research suggest that a second form of representation is not necessary
Can the debate really be resolved empirically? Proponents of parsimony can always account for the imagery results with a digital model.
But is it worth the effort?

Chapter 2: Section 2.3

You have seen examples of some theories and models in cognitive science
How should we evaluate alternative theories of human thought processes?

Example: The Endowment Effect

People place a higher value on something they already possess than on the same object if it is available for purchase. How can we explain that? Consider some theories that have been proposed:
“The pain of losing the object exceeds the joy of attaining it”
“The effect prevents intra-species conflicts that would harm the group”
“The utility function for losses is steeper than the utility function for gains”
"Potential gain and potential loss activate different regions of the brain”

Cognitive Science and Reverse Engineering

Reverse engineering: Determining design by considering function
What does the behavior accomplish? How does it do it?
(E.g., how does a manufacturing company find out what its competitors are doing?)
Reverse engineering generates an abstract task analysis, which suggests how the hardware might be realized
This leads to the idea of levels of analysis

Levels of Analysis

Marr: Vision		Stanovich: Reasoning
Computational	What purpose is served by the behavior?	Intentional	What are the person’s goals, knowledge, and beliefs?
Algorithmic	How is the purpose achieved?	Algorithmic	What cognitive processes account for these intentions?
Implementational	How is the algorithm implemented?	Biological	What biological processes operate to make this happen?

What are the critical differences between computational level analysis, algorithmic level analysis, and implementational level analysis?

We explain a particular kind of behavior using reverse engineering. Three questions:
1. What does the behavior accomplish?
2. What abstract rule does it use to do it?
3. What physical implementation of the abstract rule is employed?

The lecture discussed other versions of the level-of-analysis concept, especially Stanovich's. How are Marr's and Stanovich's descriptions alike? How are they different?

Alike:
Three levels, three separate purposes for the theory
The three levels: Purpose, Algorithm, Mechanism

Terminology (“Computational” has always been a source of confusion)
Behavior of interest: Vision and other basic processes versus reasoning and other higher order processes
Differences in the focus at each level
Differences in the focus at each level: “Purpose” at the top level has a different meaning
Marr: evolutionary adaptation
Stanovich: sentient organisms

Does it make sense to ask for a particular behavior, which level of analysis is best? Why, or why not?

Different theories serve different goals
Both classifications imply that one needs to work at all three levels

How useful are Marr’s and Stanovich's approaches to the explanation of behavior?

Consider this common dispute: “Why is Fred so obese?”
1. He doesn’t really want to lose weight
2. Food serves as a reinforcer when he’s depressed
3. His metabolism is too low
There is nothing incompatible about these explanations!