Transcript Cognitive science and the integration challenge

Chapter 4:
Cognitive science and the
integration challenge
What is cognitive science?
• Is there such a thing as cognitive science (as opposed to
the cognitive sciences)?
• Are there any basic assumptions about the mind that
unify all the different models and approaches we have
been considering?
• Does cognitive science have a distinctive methodology?
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
An early model
• Identifies connections between
different cognitive sciences,
but sheds no light on why
there need to be such
connections
• Doesn’t help us to understand
what each discipline
contributes
• Doesn’t give a very accurate
picture of contemporary
cognitive science
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
Two basic assumptions
(1) The mind is too complicated to be understood using the
resources of any discipline on its own

•
Need for interdisciplinarity
(2) The mind has to be understood at many different levels

Need for an integrated multilevel explanation
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
The example of neuroscience
Neuroscience has evolved enormously since the early days of
cognitive science
It is itself highly interdisciplinary and multilevel
Illustrates the close connection between
• levels of organization
• levels of explanation
• available scientific tools and techniques
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
Levels of explanation
A
B
C
D
E-H
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
Experimental psychology
Cognitive neuroscience
Behavioral neuroscience
Systems neuroscience
Cellular neuroscience
Molecular neuroscience
Tools and levels in neuroscience
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
Organization in psychology
• Scientific psychology is not organized vertically
in terms of levels
• Instead organized horizontally in terms of
domains
MIND AND BEHAVIOR AT THE LEVEL OF THE ORGANISM
General cognitive capacities
COGNITIVE PSYCHOLOGY
Cognition in a social context
SOCIAL PSYCHOLOGY
Individual personality and character
PERSONALITY PSYCHOLOGY
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
Non-human cognition
COMPARATIVE PSYCHOLOGY
How cognitive abilities develop
DEVELOPMENTAL PSYCHOLOGY
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
The integration challenge
•The challenge of providing an unified account of
cognition that draws upon and integrates the
whole space
• Many regions within the “space” of
cognitive science remain little studied
• The “space” is not organized by discipline
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
3 approaches to IC
Local integrations
• Examples of specific cases where cognitive
scientists have built bridges across levels of
explanation and between disciplines
Global models of integration
• Blueprints for solving the integration challenge
The mental architectures approach
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
Local integration 1
• Inspired by experiments in the psychology of reasoning
showing that people are systematically prone to very
specific types of fallacies in deductive reasoning
• Evolutionary psychologists suggest that these patterns of
error are by-products of highly specialized cognitive
mechanisms that emerged to solve adaptive problems
• massive modularity thesis
• cheater detection module
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
Local integration 1
•Solution of adaptive problems
Explains
•Emergence of dedicated cheater detection system
Explains
•Patterns of error in logical reasoning tasks
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
Local integration 2
•Cognitive scientists often treat functional neuroimaging as
if it provides direct “map” of the mind in action
•But neuroimaging directly measures blood flow (PET) or
levels of blood oxygen (fMRI) – whereas the “motor” of
cognitive activity is neuronal activity
•How do we integrate BOLD signal information with
information about neuronal activity?
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
Local integration 2
•Research question: What are the neural correlates of the
BOLD signal?
•Experiments by Logothetis explored this question by
calibrating fMRI data with microelectrode recording
• bridging levels of organization and levels of
spatial resolution
• allowed fMRI data to connect up with neural
models (and computational neuroscience)
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
Models of global integration
Two main candidates:
• Marr’s tri-level hypothesis - explicitly proposed as a way of
bridging different levels of explanation
1. computational level
2. algorithmic/representational level
3. physical level
• Models of inter-theoretic reduction derived from
philosophy of science
– analogy with unity of science hypothesis
in the physical sciences
Cognitive Science
 José Luis Bermúdez / Cambridge University Press 2010
Psychology of reasoning
• Psychologists have produced evidence that
subjects regularly contravene basic principles of
deductive logic probability theory when engaged
in
• conditional reasoning
• judgments of likelihood
Conditional reasoning rules
Modus ponens
If p then q
p
Therefore q
Modus tollens
If p then q
not-q
Therefore not-p
Affirming the consequent
If p then q
q
Therefore p
Denying antecedent
If p then q
not-p
Therefore not-q
Wason selection task
What cards need to be turned over to evaluate:
If a card has a vowel on one side then it has an
odd number on the other side
Cassava root studies (Cosmides and
Tooby)
Background about (imaginary) Pacific island:
• Only married men have facial tattoos
• Cassava roots are a highly prized delicacy and
aphrodisiac
• Molo nuts are bitter and not valued in the
community
Social exchange
Two versions of cassava root story
Descriptive: married men live on the side of the
island where cassava roots grow, while
unmarried men live on the side where the molo
nuts grow
Social exchange: only married men have the
right to eat cassava roots.
Test and results
•If a man is eating cassava root, he must have a
tattooed face
EATS
CASSAVA
ROOT
EATS
MOLO
NUTS
TATTOO
NO
TATTOO
Descriptive version
Poor performance (21%)
Social exchange version
Better performance (75%)
Cosmides and Tooby analysis
EATS
CASSAVA
ROOT
•
Benefit
EATS
MOLO
NUTS
No benefit
TATTOO
Cost paid
NO
TATTOO
Cost not paid
Social exchange version has following structure
If BENEFIT then COST
Cheater = BENEFIT without COST [i.e. p & ~ q]
Local integration 1
•Solution of adaptive problems
Explains
•Emergence of dedicated cheater detection system
Explains
•Patterns of error in logical reasoning tasks
The structure of the argument!!
•If CONDITIONAL REASONING EXPLOITS A CHEATER
DETECTION MODULE (p) then PERFORMANCE WILL
BE BETTER ON THE SOCIAL EXCHANGE VERSION (q)
•PERFORMANCE IS BETTER ON THE SOCIAL
EXCHANGE VERSION (q)
•Therefore, CONDITIONAL REASONING EXPLOITS A
CHEATER DETECTION MODULE (p)
Switched selection task
Standard social exchange selection task
• If BENEFIT (p) then COST (q)
• violation = p and not-q
Switched social exchange selection task
• If COST (p) then BENEFIT (q)
• violation = q and not-p
• Subjects typically give the logically correct answer on the standard
version, but not on the switched version
• Detecting a violation of the switched version is not the same as
detecting a counter-example to the conditional
Cosmides and Tooby analysis
EATS
CASSAVA
ROOT
•
Benefit
EATS
MOLO
NUTS
No benefit
TATTOO
Cost paid
NO
TATTOO
Cost not paid
Switched social exchange version has following structure
If COST then BENEFIT
Logically correct answers are cards 2 and 3
Cheater detection answers remain 1 and 4
Evolutionary psychology and
conditional reasoning
• evolutionary psychologists reject the idea of domain-general
reasoning skills
either mental logic or mental models
• suggest that we employ context-dependent inference rules – in
particular, rules for detecting cheaters in social exchanges
• integrate these experimental data with a model of how the mind is
organized and how it evolved
Massive modularity thesis
Gives a picture of the overall organization of the mind
• mind composed of highly specialized cognitive
modules (Darwinian modules)
• each module evolved to solve a particular
adaptive problem
• each module exploits specialized rules that are
domain-specific
• No domain-general “central cognition” or
abstract reasoning mechanism
Cheater detection module
The Cosmides/Tooby experiments seem to show
specialized skills for cheater detection
• not simply specialized skills for conditional
reasoning involving social exchanges
These experimental results are integrated with the
massive modularity hypothesis via an evolutionary
explanation of why there needs to be a cheater detection
module
• evolutionary explanation itself grounded in an
account of the evolution of altruism
The puzzle of altruistic behavior
• Cooperative behavior widespread in animal kingdom
• even in lower animals ants, termites, bees etc (individuals
fed by others etc)
• not restricted to kin
• Cooperative behavior presumably has a genetic basis
• But how did the genes coding for cooperative behavior ever get
established in the gene pool?
• natural selection seems to favor “selfish”
behavior - free riders can always exploit altruists
Modeling the evolution of cooperation
The prisoner’s dilemma is a very useful tool for modeling the
problem
• we can assume that participants are purely
selfish
• set up so that cooperation is not the dominant
strategy for
• can easily be extrapolated to multi-person
interactions (tragedy of the commons)
One-shot PD
•
Player A
•
•Player
•B
•
COOP
DEFECT
COOP
5, 5
10, 0
DEFECT
0, 10
2, 2
•Illustrates basic structure of interactions where being a free
rider is advantageous
Decision-making in a one-shot PD
• Work backwards from what the other agent might do
• Look at your options if the other agent
cooperates – it is best for you to defect
• Look at your options if the other agent defects –
it is best for you to defect
•The dominant strategy for each play is DEFECT
•But mutual defection is sub-optimal
Iterated PDs
• A backwards induction argument shows that DEFECT is
dominant when the number of plays is known
• But for modeling the evolution of cooperation the
interesting case is the indefinitely iterated PD
• opens up possibility of strategies that “punish”
other player for defecting
• and “rewarding” for cooperating
Axelrod’s computer tournament
Invited game theorists to submit strategies for iterated PD
tournament
• played strategies against each other for around
200 iterations
Highest average score came from TIT-FOR-TAT
• Start by cooperating
• Then do what the opponent did on the previous
round
TIT-FOR-TAT
• Shows how cooperative behavior might emerge in very
simple organisms
• and be maintained since, in the right conditions,
it is an evolutionarily stable strategy
• Some evidence that TIT-FOR-TAT is followed in the animal
kingdom (3-spined sticklebacks)
• Has been used to model complicated human interactions
(e.g. voting patterns in US Senate)
Back to cheater detection
• TIT-FOR-TAT (or some similar strategy, such as TITFOR-TWO-TATS) can only work if there is a reliable
mechanism for detecting cheaters. . .
• Evolutionary pressure for selection of cheater detection
module
• According to Cosmides and Tooby, this module explains
the pattern of choices made in conditional reasoning tasks
Local integration 1
•Solution of adaptive problems
Explains
•Emergence of dedicated cheater detection system
Explains
•Patterns of error in logical reasoning tasks
언어처리의 인지신경과학적 접근
• 언어처리(language processing)란?
– 정보처리적 (informational processing) 관점에서, 사람의 말과 글을
산출하고, 이해하는 과정을 지칭하는 용어
– 말(말소리)(speech, vocal output): 사람의 성대와 구강을 통해
조음되어 나오는 의미 있는 소리
– 글(text) : 말소리가 일정한 표기체계( writing system)를 통해 표현된
물리적 실체
– 말소리 또는 글을 산출하거나 이해하는 과정이 어떤 단계로 구성될
수 있으며, 그리고 그 단계들이 어떻게 유기적으로 상호작용하여
최종적인 산출이나 이해에 도달하게 되는지를 연구함.
• 예) 음소지각-음운표상/처리-단어표상/처리-문장구성/처리-담화/덩이글
표상과 처리
4
언어처리의 인지신경과학적 접근
• 인지신경과학(Cognitive Neuroscience)이란?
– 인지과학(Cognitive Science)의 중심학문
– 인지심리학:인간의 마음의 작용을 자극-반응(stimulus-response)의
조건형성(conditioning)만으로도 충분히 설명할 수 있다는 행동주의
(behaviorism) 심리학을 넘어서, 기억/주의(attention) 언어처리 등과
같은 인지과정이 마음의 중추적인 작용이라 상정하고, 이를
정보처리적인 모델로서 구현하고, 연구실에서 이루어지는 행동실험/
컴퓨터 시뮬레이션 등을 통해 이를 검증하는 학문
– 신경과학: 신경세포(neuron)의 정보전달체계를 연구하는 학문으로
작게는 단일 뉴런의 생성과 정보처리, 크게는 중추신경계의 구조적
연결과 기능을 연구하며, 20세기에 들어 중추신경계 (뇌)의 기능과
구조가 연구의 초점
– 인지심리학과 신경과학의 결합
• 인지과정을 처리하는 뇌의 기능과 구조를 연구함
• 비침습적(non-invasive) 뇌기능 연구 방법론의 발달
• 21세기는 뇌 과학의 시대
4
인지신경과학적 연구방법들
• 침습적(invasive) 방법들
Multi-unit activity, local
field potential(LFP)
4
인지신경과학적 연구방법들
• 침습적(invasive) 방법들
– George Ojemann (1989)’s direct cortical stimulation during brain
surgery.
– 간질환자의 간질부위 절개시술 전 언어영역을 찾기 위해
전기막대로 언어영역을 자극함
– 뇌 국소마취, 환자는 깨어있음
그림을 보여주고,
그림의 이름을 말하게 함
4
인지신경과학적 연구방법들
• 비침습적(non-invasive) 연구방법들
– TMS(Transcortical Magnetic Stimulation)
4
인지신경과학적 연구방법들
• 비침습적(non-invasive) 연구방법들
– EEG(Electroencephalogram), ERP(event-related potential),
– MEG(Magnetoencephalogram)
4
인지신경과학적 연구방법들
• 비침습적(non-invasive) 연구방법
– Neuroimaging 방법의 대두
– PET(positron emission topography)
4
신경과학적 연구방법들
• 비침습적(non-invasive) 연구방법
– Neuroimaging 방법의 대두
– fMRI(functional Magnetic Resonance Imaging)
4
뇌와 언어
• Paul Broca (1861), French physician
– 언어산출에 장애를 보이는 두 명의 환자의 뇌를 사후 해부하여,
좌하전두엽이 언어산출에 관여함을 주장함.
– 오른쪽 뇌의 유사부위를 다친 사람은 언어산출에는 크게 장애를
보이지 않음
– 좌하전두엽을 Broca’s region(브로카 영역)이라 일컬음.
• Carl Wernicke (1884), German psychiatrist
– 왼쪽 상/중측두엽 - 언어이해
4
실어증 환자 연구
• Broca’s aphasia:
–
–
–
언어산출의 장애(기능어(전치사, 조사), 동사 태/시제 등의 손실)
심각한 말더듬, 내용어 중심의 산출
문법구조가 복잡한 문장이해의 장애
•
•
–
Reversible sentences – picture matching)
“The white dog was chased by the brown horse.”
실험지시문을 이해
• Wernicke’s aphasia
–
–
–
언어이해의 장애
의미 없는 말의 나열(word salad), 기본적인 통사구조는 유지
실험지시문을 이해하지 못함
• Conduction aphasia(전도성 실어증)
–
–
4
Broca 영역과 Wernicke 영역을 잇는 뇌섬유다발(Arcuate Fasciculus)의 손상
자발적 산출과 이해에 크게 손상은 없으나, 따라말하기 과제에 특수한 장애를 보임
Wernicke-Geshwind model - 1965
음운표상처리
Motor
Cortex
Arcuate
Fasciculus
Angular
Gyrus
언어산출/
통사처리
음운표상
-철자표상
전환
Broca’s
Area
Wernicke’s
Area
청각적
단어 재인
5
Supramarginal
Gyrus
Auditory
Cortex
언어이해/
의미처리
시각적
단어 재인
Visual
Cortex
실어증 연구의 한계
• 뇌손상 부위의 다양성
– 같은 장애를 보이는 환자들의 뇌손상부위가 다양하며, 비슷한
뇌손상 부위를 가져도 다른 장애를 보이는 경우가 많음
– 정확한 뇌영역의 위치와 기능을 상세히 규명하기 어려움
• 뇌손상 원인의 다양성
– 물리적 외상/뇌경색/ 치매 등 다양한 원인
– 발병시기의 다양성
여러 변인들을 통제할 수 있는 인지신경과학적 실험적
방법들을 필요로 함.
5
What is the functional MRI?
• MRI(Magnetic Resonance Imaging) + a fast imaging technique
3T magnet
fMRI time-series
Statistical Parametric Map
RF Coil
gradient coil
(inside)
•
•
•
•
•
5
신경세포의 활동은 뇌의 국소적인 에너지 대사량을 증가시킴
에너지 대사에 필요한 산소는 헤모글로빈에 의해 전달됨
산소를 빼앗긴 헤모글로빈은 자기장에 영향을 주는 상자성 물질
MRI 이미지는 뇌의 구성물질이 자기장의 영향을 주는 정도를 표현함.
따라서, 신경세포 활동에 따른 국소적인 에너지 대사를 나타내는 산소소모량의
변화의 시간적 추이를 측정하는 방법
Neural activity  BOLD signal
Psychologi
cal variable
stimulatio
n
Behavioral response
Task
Task
Neuronal response
Direct measurements
Single cell recording
Local field potential
Multi-units activity
EEG(ERP), MEG
5
3
Linear transformation
BOLD signal
Hemodynamic response
fMRI, PET
뇌영상 방법을 통한 언어처리 연구의 실제
• Lee and Newman(2009), Human brain mapping
• The Effect of Presentation Paradigm on Syntactic Processing:
An Event-Related fMRI Study
– 문장 제시 방법을 조절하여 문장처리와 기억과의 관계 분석
– 문장의 통사적 난이도 조절(평서연결문 vs. 목적관계절 삽입문)
• The pilot scared the escort and broke the mirror on the closet.
• The pilot that the escort scared broke the mirror on the closet.
– 목적관계절 삽입문은 보다 많은 작업기억을 요구함. => 문장의 즉시적 처리의
어려움
– 문장 제시 방법의 차이(전체 제시, 한단어씩 제시)
5
5
문장이해 테스트 행동반응 결과
• 어려운 문장, 즉 목적관계절 삽입문을 읽고 대답할 때,
반응시간이 더욱 느리고, 더 많은 오반응을 보임.
• 문장을 한 단어씩 제시했을 때, 반응시간에서는 별다른
차이를 보이지 않았으나, 오반응률은 조건간의 차이가
훨씬 큼. => 문장 제시 방법이 문법적 처리/이해 과정에
영향을 미침.
5
문장을 읽는 과정에 나타난 문장 난이도의 효과
•
•
•
5
Broca 영역에서 통사적으로 복잡한 문장을 처리할 때 보다 많은
신경세포의 활동이 관찰됨.
그러나, 문장을 한 단어씩 보여줬을 때는, 문장의 난이도에 그리 상관없이,
전반적으로 높은 신경활성화를 보임
문장처리에 관여하는 작업 기억(Working memory)과 문장제시방법의
상호작용(interaction)을 보여줌.
문장을 읽고 난 후, 답하는 과정에 나타나는
문장 난이도의 효과
• 문장을 읽고 난 후, 답을 하는 과정에도 읽은 문장의 난이도가
여전히 영향을 미침.
• 읽은 문장의 이해정도에 따라 그 차이를 보임
• Wernicke 영역에서 문장제시 방법에 따른 상호작용을 보임
• 한 단어씩 제시했을 경우, 문장의 이해 정도의 차이가 없음.
5
의미 표상의 문제
• 우리는 어떻게 의미를 표상하고 처리할까?
• 지식 표상체계- 개념/명제/도식적 지식표상
• 의미표상의 제 문제에 대한 인지신경과학적 접근
–
–
–
–
5
구체적 사물에 대한 의미 (inferior temporal visual pathway)
행위에 대한 의미 (motor cortex)
추상적 개념에 대한 의미(inferior parietal lobule)
=> 우리가 어떻게 대상개념을 습득하고 상호작용하는가에 의해
결정되는 것 같다.
Semantic memory in the brain
Part1: objects
Neural representation for object concepts
(Animal vs. Tool) adapted from (Chao & Martin,
Nature, 2001)
Red – Animal, Blue- Tool
Semantic processing for object concepts (Martin, Annu. Rev. Psy. 2007)
FFA(fusiform face area) – Kanwisher et al.(1997), McCarthy et al.(1997)
PPA(parahippocampal place area)- Epstein and Kanwisher(1998)
6
Semantic memory in the brain
Part2: actions
Neural activation for
6 Action Words
Somatotopic representation for action
concepts
adapted from (Hauk et al., Neuron, 2004)
단어의 의미에 기초한 뇌 활성화 예측
• Predicting Human Brain Activity Associated
with the Meanings of Nouns (Science, 2008)
6
마치며…
• 뇌와 마음
– 뇌에 마음이/언어가 있는가?
– 뇌는 소중한 우리 신체의 일부, 손/발 과 마찬가지로..
– 뇌는 우리의 정신적/육체적 작용의 학습과 발달로 끊임없이
변해가는 기관.
– 따라서, 정신적/육체적 활동의 변화를 잘 나타내는 신체기관으로,
이를 연구함으로 인해 우리 마음의 작용을 보다 과학적으로
연구할 수 있는 토대를 제공함.
– 여러분은 무엇을 연구할 것인가?
6