How Can the Human Mind Occur in the Physical Universe?

How Can the Human Mind Occur in the Physical Universe?

John R. Anderson

User Interface Lab

Contents

1. Newell’s ultimate scientific question

2. What is a cognitive architecture?

3. Alternatives to cognitive architectures

4. ACT-R: a cognitive architecture

5. Symbol vs. connections in a CA

2 / 26

User Interface Lab

1. Newell’s Ultimate Scientific Questions (1/2)

Allen Newell (March 19, 1927 ~ July 19, 1992)

Ultimate Scientific Questions• Why does the universe exist?• When did it start?• What’s the nature of life?

for Newell’s• How can the human mind occur in the physical

universe? ※ this question leads him down to worry about the architecture

3 / 26

Last lecture (Dec 4, 1991)

“Desires and Diversions”

User Interface Lab

1. Newell’s Ultimate Scientific Questions (2/2)

Purpose of this book is to report on some of the progress that has come from

taking a variety of perspectives, including biological

Answer would be like : cognitive architecture

Purpose this chapter What is cognitive architecture? How the idea came to be What the (failed) alternatives are Introduce the cognitive architecture

4 / 26

User Interface Lab

2. What is a Cognitive Architecture? (1/4)

Cognitive Architecture

Architect is concerned with how the structure achieves the function.

structure (domain of the builder) function (domain of the dweller) ☞ Architecture is the art of specifying the structure of the building at the level of abstraction sufficient to assure that the builder will achieve the functions desired by the user.

Newell (1971) introduced Cognitive Architecture through an analogy to Computer Architecture

Fred Brooks (1962) introduced into computer Science through an analogy to the architecture of buildings.

Architecture of buildings

Architecture Computer Science Cognitive Science

5 / 26

User Interface Lab


Brooks (in Planning a Computer System) computer architecture is the art of determining of user

needs and d then designing to meet those need ☞ Brooks is using “architecture” to mean the activity of design

Definition (cognitive architecture) Newell (1990)

☞ the fixed (or slowly varying) structure that forms the framework for the immediate process of cognitive performance and learning.

Pylyshyn (1984)☞ the functional architecture includes the basic opera-tions provided by the biological substrate, say, for storing and re-trieving symbols, comparing them, treating them differently.

Anderson (1983) ☞ a theory of the basic principles of operation built into the cognitive system.

6 / 26

User Interface Lab


Structure Building’s architecture : physical components Cognitive architecture..: do not mention the brain Function Building’s architecture : habitation Cognitive architecture..: cognition

• Functional shift : activity of another → its own activity ☞ except for this shift, there is still the same S-F rela-tionship; function of the structure is to enable the behavior.

Agent (dweller)

Agent (structure)

7 / 26

User Interface Lab


Before the idea of CA emerged, a scientist has two op-tions;

either focus on structure and get lost (endless details of the brain)

or focus on function and get lost (endless details of be-havior) ☞ CA reflects the relationship between S and F rather than focusing d on either individually

Definition (for the purpose of this book)Cognitive Architecture is a specification of the structure of the brain at a level of abstraction that explains how it achieves the function of the mind

Function of the mind : Can be roughly interpreted as re-ferring to d human cognition in all of its complexity

8 / 26

User Interface Lab

3. Alternatives to Cognitive Architecture (1/8)

The type of architectural program requires payingd attention to three things; Brain, Mind, Architectural abstraction

This chapter examines three of the more prominentd Instances of such shortcuts, Success :

• discuss what they can accomplish Demerit :

• Note Where they fall short of being able to answer Newell’s question.

Problem :• What their problems are

Archi-tecturalAbstrac-

tion

9 / 26

User Interface Lab

Shortcut 1. Classic Information-Processing Psychology: Ignore the Brain

Success Info-Processing Psychology was very successful

during1960s~1970s• inspect human brain → neural explanation is too

complex• so, We need a level of analysis that is more abstract

for example : Sternberg task & model Demerit “computer-inspired ” model of discrete serial search Problem ignore the brain (structure) is like a specification of a buildings architecture that

ignored what the building is made of

10 / 26


User Interface Lab


Saul Sternberg’(1966) task & model of it

information processing stage

• comparison time : 35~40 msec Sternberg reached for the computer metaphor

“when the scanner is being operated by the central process it delivers memory representations to the comparator. If and when a match occurs a signal is delivered to the match register”

See a small num-ber of digit“3 9 7”

Keep in mindAnswer whether a particular digit is in

this memory set

11 / 26

User Interface Lab


Connectionism arose in the 1980s bolstered Anderson’s general claim

• information processing between brain and computer

Neural imaging arose in the 1990s showed the importance of understanding the brain as

the structure underlying cognition.

showed where cognition played out in the brain.

• Parallel but slow• Continuous (Neurons in the

Brain)Brain

• Sequential and rapid• DiscreteComputer

12 / 26

User Interface Lab


Shortcut 2. Eliminative Connectionism: Ignore the Mind Success notable success during 1980s~1990s abstract description of the computational properties of

the brain • “neurally inspired” computation

for example : Rumelhart and McClelland’(1986) past-tense model

Demerit is not concerned with how the system might be

organized to achieve functional cognition

Problem ignores mental function (Mind) as a constraint and just

provides an abstract characterization of brain structure

all we have to do is pay attention to the brain; just describe what is happening in the brain at some level of abstraction.

13 / 26

User Interface Lab


Rumelhart and McClelland’(1986) past-tense model children, with irregular past tense sing : sang → singed → sang : conventional wisdom

•correct irregulars, over generalize, get it right past-tense model

simulating a neural network : learned the past tenses of verbs

☞ one can understand function by just studying structure

sleight of hand becomes apparentThis is not a common human behavior

14 / 26

User Interface Lab


Shortcut 3. Rational Analysis: Ignore the Architecture

Success RA (e.g., vision, memory, categorization) have characterized

features of the environment that all primates experience Demerit

rather focus on architecture as the key abstraction, focus on adaptation to the environment ☞ rational analysis (Anderson, 1990) ☞ Anderson’s application of this approach was Bayesian

Problem Human mind is not just the sum of core competences such

as memory, or categorization, or reasoning

15 / 26

User Interface Lab


Bayesian approach a set of prior constraints about the nature of the world given various experience, one can calculate the

conditional probability given the input, one can calculate the posterior

probabilities from the priors and conditional probabilities.

after making this calculation, one engages in Bayesian decision making and take the action that optimizes our expected utilities

☞ the world makes on our memory (Fig 1.4. e-mail message) ※ indicates that time since a memory was last used is an important determinant of whether the memory will be needed now

16 / 26

User Interface Lab

4. ACT-R: a Cognitive Architecture (1/4)

Goal of this book is to use one architecture (ACT-R) to try to convey what we

have a learned about human mind

ACT-R’s Modular Organization visual module hold the representation (3X-5=7) problem state module (imaginal module) hold a current mental rep’ of the problem (3X=12) control module (goal module) keeps track of one’s current intentions declarative module retrieves critical info’ form memory (7+5=12) manual module programs the output (X=4)☞ each of these modules is associated with specific brain regions

※ ACT-R contains elaborate theories about the internal processes of these modules

17 / 26

Fig 1.5. The interconnections among modules in ACT–R 5.0

User Interface Lab 18 / 26


ACT-R’s Modular Organization production system (sixth module : central procedural

module) can recognize patterns of info’ in the buffers and

respond by sending requests to the modules

these recognize-act tendencies are characterized by production rules

production rule If the goal is to solve an equation, and the equation is of the form “expression – num1= num2,” Then write “expression = num 2 + num1,”

Experiment : children 11~14 years of age three classes of equations on a computer: 0-step: e.g., 1X + 0 = 4 1-step: e.g., 3X + 0 = 12, 1X + 8 = 12 2-step: e.g., 7X + 1 = 29

1818

Fig 1.6. Mean solution times (and predictions of the ACT–R model) for the three types of equations as a function of delay.

User Interface Lab


Brain Imaging Data and the Problem of Identifiability children’s 5 brain regions were scanned : Fig 1.8 they are associated with specific modules in the ACT-R

theory

Predicting the BOLD Response in Different Brain Regions x-axis : time (from the onset of the trial) left graph : effect of number of operations averaging

over days right graph : effect of days averaging over operations

• response shifts a little forward in time from day 1 to day 5, reflecting the speed increase

19 / 26

User Interface Lab


Summary

1. unlike the classic info-processing approach, the architecture is directly concerned with data about

the brain.

2. unlike eliminative connectionism, an architectural approach also focuses on how a fully

functioning system can be achieved.

3. unlike the rational approach and some connectionist approaches,

ACT-R does not ignore issues about how the components of the architecture are integrated.

20 / 26

User Interface Lab

5. Symbols Vs. Connections in a CA (1/6)

Debate notorious debate between symbolic and connectionist

architecture there is no consensus about what role symbols play in an

explanation of mind ※ “+” indicate an explanatory role, “-” non explanatory role

1.+symbols, -connections: transformation of the structural properties of symbolic

representations unimportant : the physical processes that realize these

symbols2. - Symbols, +Connections: this position is called eliminative connectionism• it seeks to eliminate symbols in the explanation of

cognition it views symbols much like elements in explicitly stated

rules• “if the verb ends in d or t, add ed”

21 / 26

User Interface Lab


3. +Symbols, +Connections: both play an important explanatory role

• Integrated Connectionist/Symbolic(ICS) architecture

4. - Symbols, - Connections: reject both architecture and offer other explanatory

devices• Functionalism, some varieties of Behaviorism• situated cognition: explanation resides in what is

outside the human

※ Because there is not agreement about what symbols mean, these debates are a waste of time

22 / 26

User Interface Lab


Symbolic-Subsymbolic Distinction

symbolic level in ACT-R is an abstract characterization of how brain structures

encode knowledge. subsymbolic level is an abstract characterization of the role of neural

computation in making that knowledge available. Newell (1990) identifies the critical role of symbols

symbol provide distal access to knowledge access• information must be brought from other locations

this is exactly what they do in ACT-R; Question what info’ will be brought and how quickly that info’ will

appear this is what the subsymbolic level is about 23 / 26

User Interface Lab


Symbolic-Subsymbolic Distinction in the Declarative Module Sugar factory task (Fig 1.9)

Chunks (symbolic level) ACT-R has networks of knowledge encoded in what we

call chunks chunks have activations at the subsymbolic level Activations (subsymbolic level) most active chunk will be the one retrieved Its activation value will be determined by computations

that attempt to abstract the impact of neural Hebbian-like learning and spread of activation among neurons.

24 / 26

User Interface Lab


Symbolic-Subsymbolic Distinction in the Procedural Module PM consists of production rules illustration of a production rule in ACT-R (Fig 1.10)

general pattern • information location

☞ symbolic level

Multiple production rules applied situation production have utilities and production with highest

utility is chosen ☞ subsymbolic level 25 / 26

User Interface Lab


Final Reflections on the Symbolic-Subsymbolic Distinction

confusion Nothing in the production rule in fig 1.10 is different

from the pattern- matching capabilities of standard connectionist networks. Actual code looks like cognitive science stereotype of a

symbol as a piece of text

• symbol for the simulation program, not the symbols of the ACT-R

architecture level of description

choosing best level is a strategy decision ACT-R : higher level processes such as equation solving gap is smaller in the case of ACT-R (from neurons and

brain process) the same level of description might not be best for all

applications. Connectionist model : perceptual processing

26 / 26

How Can the Human Mind Occur in the Physical Universe?

Documents

Transcript of How Can the Human Mind Occur in the Physical Universe?