Summary Foundations of Cognitive Science
            (c) 2001-01-14 Fabian M. Suchanek
     http://www.mpi-inf.mpg.de/~suchanek/personal/texts/summaries/foc.txt


This is a summary of the course "Foundations of Cognitive Science"
held by Prof. Peter Bosch and Prof. Franz Schmalhofer in the
WS 2000 at the University of Osnabrueck. It is partially based on
"Summary of CogSci" (c) 2000-12 Serap Avci.

By reading the following text, you accept that the author does not
accept any responsibility for the correctness or completeness of
this text. If you have any corrections or remarks, please
send me a mail. This is the only way to make the publication of this 
summary useful for me, too.

My e-mail address is f.m.suchanek@zweb.de, but the letter 'z' has to 
be removed from the address.


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    Cognitive Science
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Information processing:
 receiving, storing, retrieving, transfroming and transmitting data

CogSci:
 studies cognitive functions from an information processing point of
 view.
 It is a cocktail of its mother disciplines psychology, linguistics,
 computer science, philosophy and neuroscience plus the "Cognition as
 Computation" paradigm.

Cognition as computation:
 the brain is seen as a processor with a memory, in addition to
 receptors and effectors.
 Cognition is representable by symbols which are manipulated.
 Formals symbol manipulation is carried out w/out thought or intention.
 The result of a calculation only depends on the symbols, not on their
  meaning.

Computation:
 Computation manipulates formal symbols
 Symbols are representational
 Symbol manipulation is purely syntactic
 Symbol manipulation is semantically invariant
 Explicitness about internal, symbolic mechanisms
 Representational properties of symbols are intentional and are
  external to the machine

Pattern of natural sciences:
 A causes B
 Problem: humans do a for obtaining b, so b is the reason for a

3 levels of description:
 1. knowledge level
    sentences about the real world
 2. formal level
    the formal representation of the knowledge level by symbols
 2. physical level
    the formal representation must be implemented physically

Semantic mapping:
 Both knowledge representation and formal representation are mapped
  on the same state in the real world.


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    Testing the Short Term Memory
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Quality criteria of empirical research:
 * Objectivity
 * Reliability (can be repeated)
 * Validity (relevance)

The Stroop-Task:
 Task: names of colors are written in different colors, the task is
 to name the colors of the writing.
 Result: It takes longer to name the color of the writing when the
 color is different from the color name.
 Conclusion: Different speed in reading and naming colors caused by
 different training.

Conclusions of the Stroop effect:
 Cognition is
 * non-conscious
 * automatic
 * parallel

Further interpretation of the Stroop effect:
 Parts of the system are informationally encapsulated: The visual
 system and the language system return different information in the
 Stroop-Task which leads to an inhibition.

Serial position effect:
 Task: 14 words are shown each for 1 second, then they are to be written
 down.
 Result: The first words of the list are remembered well, the words
 in the middle are remembered badly and the words at the end are
 remembered best.
 Conclusion: Words go from short term memory to long term memory.
 The good rate at the beginning is called primacy effect: The words
 have a better chance of getting stored in the long term memory due to
 better concentration at the beginning. The words at the end are still
 in short term memory and can thus be recalled very easily.
 Repeating this task with clearing the short term memory afterwards
 (by counting backwards e.g.) results in a equally bad rate at the end.

Chunk theory:
 7 +-2 items can be kept in short term memory. These items are
 cognitive units, called chunks. Chunks seem to be independent from
 technical information content (in bits). Chunks can be numbers,
 words, or letters.

Short term memory (STM) and Working memory (WM):
 The name "STM" stresses the durability of memory contents
 The name "WM" stresses the fact that cognitive processes work
 with these memory contents.
 Often STM and WM are used synonymously.
 STM can be suppressed (by the counting task)
 Rehearsal maintains information in the STM
 The STM has a limited capacity (Chunk theory)

Finding subsystems:
 One tries to identify independent subsystems of human cognition.
 A model is best if the systems are independent, so
 * maximize interaction in a system
 * minimize  interaction between systems

Memory model according to Atkinson & Shiffrin (1971):
 Environmental input -> sensory registers -> STM <-> LTM
                                              '--> Output

Word length experiment:
 Task: Words with different lengths are to be remembered.
 Result: The more syllables a word has the more difficult it is to
 remember.
 Problem: This effect cannot be explained by the chunk theory.
 Conclusion: Material is encoded verbally

The phonological loop (PL):
 (=Audiobuffer) we can keep 1,3 s worth of materials rehearsed in the
 phonological loop. The speed of verbal rehearsal is very close to the
 speed of spoken speech.

The Visuo-spatial sketch pad (VSSP):
 (=Videobuffer) a picture can be kept in the VSSP for a short time.

Working memory model:
 PL    <->    STM    <->    VSSP
 ^                ^
 auditory  phonological  central    visual    visual
  sensory   rehearsal   executive   rehearsal   sens.
   input                input

Secondary task technique:
 Putting heavy load on one subsystem to clear it without influencing
 the other subsystems. The PL can be cleared without influencing the
 VSSP by comparing the pitch of two tones. The VSSP can be cleared
 without influencing the PL by comparing the brightness of two
 surfaces.


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    Testing the Long Term Memory
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Memory tasks:
 The recall task: Reproducing what you learned - either with a cue
 ("cued recall") or without ("free recall").
 The recognition task: Remembering that something has been presented
 before.
 Result: In general, recogniton is easier that recall.
 Conclusion: LTM contents are addressed via cues. In a recall-task,
 the cue has first to be found while in a recognition task, it it
 given. The data is then found by "associative retrieval".

Flashbulb memories:
 ...are memories on emotional events. There is a high confidence on
 these memories, but this confidence is wrong.

Reconstructiveness:
 Memory is reconstructive: It generates something that is plausible.

Category experiment:
 Task: Participants are given a mixed list of 60 words and asked to
 recall them and write them down.
 Result: Participants recall words in categories.
 Conclusion: People categorize the world.

Categorizing the world:
 We distinguish between
 * Concepts: The categories into which we classify objects (classes).
   Concepts can be superordinated or subordinated to other classes
   (derived from other classes).
 * Instances: Concrete existing objects

Classical concept theory:
 * Concepts are arbitrary: Nothing in our nervous system or in the
   world determines how we must slice up our observation.
 * Concepts are not fuzzy: All instances represent the concept equally
   well.
 * Concepts are defined: All instances share the defining attributes.

Prototype hypothesis:
 Concepts are build around a central instance (prototype), e.g.
 "robin" for bird.

Basic objects in natural categories:
 * superordinate, e.g. instrument
 * basic level: flute
 * subordinate: wooden flute

Semantic net:
 Memory can be seen as a net of concepts which are subordinate to other
 concepts (shown by a net with "is a" arrows). All concepts have
 properties (shown by arrows labelled "property" from a concept to a
 property). Reasoning is following the arrows in the net.

Semantic priming experiment:
 Task: A word is given. Then a string is given. The task is to
 determine whether the this string is a word (lexical decision task).
 The time it took to find out the string is a word is measured.
 Result: If the first word and the second word are semantically
 related, the time needed is shorter. For polysemes, this works with
 all meanings.
 Conclusion: Conceptual network theory

Conceptual network theory:
 Conceptual knowledge is organized in a network structure. The prime
 activates a concept node. Activation spreads over the network. The
 preactivated target is recognized faster.

Association:
 ...can be direct or indirect via a superordinate concept.

Recognition test with lures:
 Task: A list of words is given and the participants are later to tell
 whether specific words have been on the list (recognition).
 Result: Participants say words were on the list which were not on the
 list -- only semantically related words (lures) were on the list.
 Conclusion: Memory can trick us and does not remember exactly.


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    Linguistics
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Natural language (NL):
 ...is the most natural way to communicate and store information.
 Language seems to be everywhere in the memory model.

Innateness hypothesis:
 Observations: Everyone can learn a language without tuition. Language
 cannot be learned via induction. It does not help to correct children
 learning to speak.
 Conclusion: There is an innate universal grammar that requires very
 little information (parameters) to get language acquisition started.

Linguistic competence:
 ...consists of a grammar which defines an infinite set of well-formed
 expressions and their relation to meaning. Linguistic competence is
 tacit knowledge: You have it, you use it but you can't make it
 explicit.

Ambiguity:
 ...means having multiple meanings.
 Lexical ambiguity: "I'm going to the bank."
 Structural ambiguity: "They are visiting relatives."

Competence vs. performance:
 Competence is the ability to speak correctly (the knowledge of the
 grammar) while perfomance is the quality of the actual output.

Center-embedded relative clauses:
 ...put a considerable burden on the WM because unfinished clauses
 must be kept in memory until they are completed by the corresponding
 verb.

Minimal attachment strategy:
 Complete a constituent as soon as you can. Don't add new constituents
 beyond necessity. This strategy leads to the Gardenpath-effect
 (described by the well-known phrase with the horse).

The importance of a frame:
 If there is no frame (szenario) to given sentences, every sentence
 must be remembered verbatim. When there is a frame, single items fall
 into place, the listener remembers the gist of the story independent
 of the verbal form.

Boundaries in sentences:
 Speech pauses occur at phrase junctures, they form the natural units
 of sentences. Intuition concerning boundaries (as found out by
 correctiones) are the same for all people.

Units of substitutions, Phrase Structure Tree, Phrase Maker,
Constituent structure:
 Grammar:
  Sentence (S) -> Noun Phrase + Verb Phrase
  Noun Phrase (NP) -> (Determiner+) (Adjective+) Noun
  Verb Phrase (VP) -> Verb (+ Noun Phrase) (+Complement Phrase)
  Complement Phrase (CP) -> "that" + Sentence
 Vocabulatory (terminal elements):
  Noun -> { John, ball,... }
  Verb -> { sleeps,... }
  ...

Tests for constituent structure:
 * Substitution
 * Question formation
 * Conjunction of identical constituents
 * Pseudo-clefting ("It is _Peter_ who arrived late.")

Meaning:
 The meaning of a sentence is the conditions under which the sentence
 is true. The meaning of a word is the contribution it makes to the
 sentence.

Understanding:
 ...is prior to deciding about the truth of a sentence.

Deciding about the truth of a sentence:
 * a posteriori, empirically, contingent upon experience
 * a priori (true in all possible worlds)
   * logical truth, mathematical truth
   * analytical truth, i.e. meaning + logical truth

Set theory:
 Suppose Noun Phrases refer to individuals and Verb phrases to sets.
 Then "Franz is a professor." would be true if the individual referred
 to as "Franz" is an element of the set of professors. To differentiate
 between the name "Franz" and the individual referred to, we write
 [[Franz]]. We write [[professor]] to express we mean the set of
 professors. This set is called the extensional meaning of the word
 "professor". Names don't have meanings. Now the above sentence
 could be written as: [[Franz]] E [[professor]].

Intension: The inherent sense of a concept.

Concepts are truth functions:
 ...from a domain for which the function is defined to a range of
 values which is 0 and 1: F a = {0,1}. It returns 1 (true) if a is
 in the extension of F. F can be defined by a rule or by a table of
 values (if the domain is finite). Concepts are often written like
 "_is <anything>".

Relation of concepts:
 Concepts are logically related in the same way that sets are related.
 * A includes B: A > B
 * A is disjunct with B: A /\ B = {}
 * A and B overlap: A /\ B <> {}

Domain:
 Concepts are restricted to domains. E.g. "_is red" can only be
 applied to material objects.

Vagueness:
 Some concepts are vague (like _isread and _isorange), incompletely
 defined or underspecified.

Semantic anomaly:
 Sentences without a meaning due to inherent contradiction.

Principle of compositionality:
 The meaning of complex expressions can be computed by the meaning of
 its parts.

Semantic terminology:
 Synonym: Meaning is identical
 Homonym: Meanings are different, but look the same
          (homphones & homographs)
 Polysem: Has different (though related) meanings
 Antonym: having the opposite meaning, often binary (dead, alive),
          but sometimes also scalar (small, medium, large)
 a is a hyponym of b: a includes b
 a is a hyperonym of b: a is included in b

Adjectives as sets:
 Adjectives can be
 * intersective: Friendly doctors = those who are friendly and doctors
 * subsective: Good doctors = those doctors who are good
 * non inter- and non-subsective: former doctors


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    Artificial Intelligence
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Intelligence:
 Intelligence is the faculty of understanding.

Knowledge:
 Familiarity with facts [, truth or principles as from study or
 investigation].

Artificial Intelligence (AI):
 Systems that think/act like humans/rationally.

George Boole:
 Investigated the laws of logic and probabilities.

Alan Turing:
 Put forward the idea that a computer could be programmed so as to
 exhibit intelligent behavior.

John McCarthy and Marvin Minsky:
 Every aspect of learning and every other feature of intelligence can
 in principle be so precisely described that a machine can be made to
 simulate it.

Architecture of an expert/knowledge system:
 expert & knowledge
 engineer        user
  |        \/
  V        /\
 knowledge acquisition    user interface
  |        \/ problems
  V        /\ results
 knowledge base (re-    search system
 presentation system)  <->  (inference system)

 Expert system reasoning has a narrow, but deep field of expertise.

MYCIN:
 Expert system for bacterial infection diagnosis
 Rule-based approach: If-Then production rules (IF A & B THEN C).
 Backward chaining: MYCIN tries to determine whether C by
 determining whether A & B, working backwards from the possible
 conclusions to the facts.

Problems with first generation systems:
 * narrow expertise
 * disgraceful breakdown at the limits of its expertise
 * no knowledge about its competence or incompetence
 * Heuristic programming approach
 * Maintenance problem
 * no possibilities for knowledge evolution

STRIPS:
 Script language for robots, consists of
 * actions (e.g. GOTO(dx))
   describing the faculties of the robot
 * its preconditions
   (e.g. TYPE(dx,DOOR), there exist rx and ry such that
   ((INROOM(ROBOT,rx) and CONNECT (dx,rx,ry)))
   describing the conditions for an action
 * its deletions (e.g. AT(ROBOT,$1,$2), NEXTTO(ROBOT,$1))
   describing the negative consequences of an action
 * its additions (e.g. NEXTTO(ROBOT,dx))
   describing the positive consequences of an action
 Frame axiom: Everything remains unchanged except for the changes
 described in the ADD & DEL lists.
 The STRIPS notation can be used for theorem proving.

Knowledge level hypothesis (Newell 1982): (?)
 There exists a distinct computer system level, lying immediately over
 the symbol level, which is characterized by knowledge as the medium
 and the principle of rationality as the law of behavior.

Principle of rationality:
 If an agent has knowledge that one of its actions will lead to one of
 its goals, then the agent will select this action.

FMS: "Lecture 8, Sheets 27 & 28 & 30" (?) missing

Development cycle (?) for knowledge-based systems:
 data -> Conceptual model -> Design model -> Implemented system

Modern knowledge systems:
 * are no longer stand-alone systems
 * interact, negotiate and communicate in multi-agent systems
 * are accessible over the internet
 * represent and reason with common sense knowledge
 * allow its users to be cooperatively involved in solving problems and
   evolving knowledge

Problem solvers:
 ...need:
 * ability to act
 * goal state
 * given state
 * set of operators
 * knowledge about the relation of operators to goal state

Formalizing problem solving:
 Problem solving requires
 * list of actions a problem solver can perform
   A={a1,a2,...}
 * set of environments the problem solver can be in
   S={s1,s2,...}
 * a set of sequences of elements of S describes episodes/histories
   S*={(s2,s8),(s9,s1,s3),...}
 A problem solver can now be described by a function K (knowledge)
 which maps sequences of environment states to actions:
   K:S*->A
 It performs actions on basis of its history.
 The consequences of an action are described by a function which maps
 the current environment and the current action to a set of
 environments which might result:
   E:S x A -> P(S) (where P(S) is the power set of S)

General Problem Solver  (GPS):
 Strategy of solving problems
 Requirements:
 * current state
 * goal state
 * set of operators
 * operator difference table
 Algorithm:
 * select the operator that reduces the largest difference
 * examine whether operator is applicable
   IF yes, apply it, else generate the subgoal of accomplishing the
   state where the operator is applicable.
 * proceed recursively
 GPS uses means ends heuristic, forward chaining.
 GPS does not perform very well, it may get into states where the
 apparent distance to the goal is short while the real distance is
 large.

The A* algorithm:
 A* finds optimal solution and reduces search cost since it does not
 need a compete search the whole search space/tree. It performs a
 "best first search": It first expands in depth and then in breadth.
 Requirements:
 * approximate information about the ressources needed to perform an
   action. This app. info must be optimistic, i.e. more ressources
   are needed in reality.
 * detailed information about the ressources nedded for an action.
   Det. info is more expensive to gain and will only be used when
   necessary.
 Algorithm:
 1. Find out the seemingly cheapest way to get to the goal state
    according to the app. info
 2. For the first segment, replace the app. info a by the detailed
    info d
 3. The cheapest path according to the new information is determined
 4. Steps 2 and 3 are repeated until a detailed estimate is obtained
    all the way for the goal state and this detailed estimate is the
    shortestof all calculated estimates.
 The break condition of step 4 occurs  without having calculated the
 detailed info for all possible ways, since in case a detailed info
 is cheaper than an app. info, the det. info way is the best:
 Calculating the det. info for the app. info way would result in an
 even more expensive result for this way.
 Technically, A* envolves a heuristic evaluation function h(i), which
 is for each possible node i of the search space the sum of the
 det. info up to the current node d(i) and the app. info for getting
 frome there to the goal state, a(i): h(i)=d(i)+a(i).
 In each step, i is increased, d(i) forms a larger part of the sum and
 a(i) a smaller part.

Human problem solving:
 Try out all possible paths to the goal and measure the ressources.
 Take the cheapest way the next time.

Computer problem solving:
 Calculate all possible paths, estimate the ressources for all paths.
 Take the cheapest.

Machine Learning:
 symbolic systems:
 * Classification (ID3)
 * concept learning (Induction)
 * explanation-based learning (EBG/EBL)
 subsymbolic learning:
 * back propagation

Classification problem:
 Objects, their properties and their class ("+" or "-") are given.
 The system has to find out whether a new object is positive or
 negative.

Learning by induction:
 In order to solve a classification problem, the system needs
 generalisation rules, such as:
 * constants can be replaced by variables
 * irrelevant features can be taken out of consideration
 * generalisation from two examples to an interval
 * going up 1 hierarchy level
 * using negation (e.g. "no square" instead of "circle")
 The so-called Star-algorithm then runs as follows: (?)
 * select a positive example (the "seed")
 * apply generalisation rules to exclude all negatives
   (results in "star")
 * simplify description (prefer conjunctive rules to disjunctive
   rules)
 * test the description for other positives
 * loop
 The learned concept is affected by representational and search biases.

Explanation-based Generalisation (EBG/EBL): (?)
 Used for example in a Safe-to-stack-problem (Kloetzchenwelt).
 Needs:
 * target concept definition, a description of the concept to be
   learned
 * training example, an example of the target concept
 * domain theory, a set of rules and facts to explain how the example
   satisfies the target concept
 * operationality criterion, specifying the form in which the learned
   concept must be expressed.
 EBG then tries to prove why the training example satisfies the
 target concept with the help of the domain theory.


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    Natural language processing
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Natural language processing (NLP):
 1. Speech recognition (mapping sppech -> text)
 2. Natural language understanding
        <planning how to react>
 3. Natural language generation
 4. Speech synthesis (mapping text -> speech)

Style correction:
 Grammar correction in text processing programs.

Concepts:
 Concepts can be represented by different words (e.g. synonyms).

Classification:
 The task of putting data into different classes.

Clustering:
 The task to find n classes of the same size so that all data is spread
 equally over these classes.

Abstracting/Summarization:
 The task of writing a summary of a text. In order to do so, the
 information is first classified.

Usage of NLP:
 * information access
 * machine translation
 * human-machine-interfaces

Information access:
 Concerning storage and seeking of data, we distinguish
 * information retrieval: Just getting out the data of the storage
   medium like it was stored
 * data mining: Getting new information out of the storage medium
   which was not stored explicitly.

Machine translation:
 input is analyzed -> semantic represenation -> generation in goal
 language -> output

Problems of machine translation:
 Machine translation is up to now far from perfect. This is due to
 structural and lexical ambiguitiy in input sentences (the system
 does still not understand the sense so that it could distinguish
 different meanings of words). Furthermore, a machine only has
 a finite dictionary while language has a nearly infinite set of words
 (words can be invented ad hoc and language varies).

Problems of multiple language systems:
 If a system shall translate n languages, it must have
 an analyzer, a transferrer and a generator for each possible pair
 of languages: n*(n-1)=n^2-n. In contrast, the human mind seems to
 translate any input to an "interlingua" first and after that translate
 this internal representation to the goal language. The task is to
 find this inner representation in order be able to teach this
 technique to computers.

Human-Machine-Interface:
 A Human-Machine interface should be able to recognize the wishes of
 its user (plan recognition). It therefore needs certain "interaction
 capabilities" such as e.g. user modeling (building up a personal
 profile for each user). Problem: When the system misunderstands the
 user or does not understand him at all, so-called "clarification
 dialogues" have to take place -- costing time and thus money.

Two views of human NLP:
 * grammatical parsing view:
   ...assumes that humans understand spoken sentences by analyzing
   their grammar (find subject, find verb etc.). It is evident that
   the human mind needs grammar (e.g. for differentiating "Jon hit Bob"
   and "Bob hit Jon").
 * integrated knowledge view:
   ...assumes that humans guess the content of a sentence by the
   content of single words (like we do when we hear somebody speaking
   in a noisy environment).

Winograd's block word test:
 Task: Execute the (ambiguos) task "Put the red cube on the block
       in the box". Its unclear whether "the red cube on the block"
       is a single item or "the block in the box" is a single item.
 Result: If only one reference succeeds, this interpretation is chosen.

Modular syntactic procession test:
 Task: Press a button as soon as you hear the word "guitar". The
 following text is read out: "The crowd was waiting eagerly. The young
 man took the guitar and..." where "took" is randomly replaced by
 more or less sensible words.
 Result: The less sensible the substitution was (e.g. "ate"), the
 longer it took people to press the button after the word "guitar".
 Conclusion: The subject cannot listen without trying to understand.

In brief:
 The interaction of grammatical knowledge and semantic knowledge in
 discourse understanding is still only poorly understood.


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    Philosophy of Cognition
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Philosophy of Cognition:
 ...could be connected with different domains:
 * Philosophy of science
 * Ethics
 * Philosophy of language
 * Philosophy of mind
 * ...

Linguistic philosophy:
 ...does not ask what it means e.g. "to believe" but what it means if
 somebody _says_ "I believe". In this example, this statement could be
 a clue about the person's future behavior.

Descartes' view:
 Differentiating between "signifiants" and "signifiés": Symbols and
 what they represent. The symbols are arbitrary and there is no
 intristic connection between a symbol and the thing it represents.
 Mind and body are two different things and none of them causes the
 other (dualism, the "ghost-and-machine-dogma" with a private soul
 and a public body).

Hobbes' view:
 * Thinking can be seen as computation.
 * Objects outside press on our senses, this pressure is passed
   through the nerves to our heart, causes counterpressure and thus
   an idea of the thing (materialism: objects exist, ideas are
   secondary)

Barkley's view:
 There are two kinds of things in human minds: First, what we perceive
 and second what is produced by our mind. Sense impressions are
 associated with something and given a name. Thus, we can just talk
 about something we perceived (esse est percipi -- to be is to be
 perceived). Objects themselves are incomprehensible, the ideas of
 these objects are essential (idealism).

Ryle's view:
 Dualism is a category mistake: Mind and body are of different
 categories and cannot be grouped together.

Fodor and Putnam's view:
 The complicated thing "mind" is composed of simple physical processes
 (functionalism).

CogSci's view:
 Mixture of the ideas of representation (Descartes), computation
 (Hobbes) and functionalism (Fodor and Putnam).

Problem of intention:
 How can we make sure that our representation of something is really
 the representation of this very thing? CogSci tries to find a
 connection between representations and the real world.


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    Neurobiological Principles
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

see also: "Summary on Introduction to Neurobiology"
          (c) 2001 Tobykenobi & Sven
          at www.coxi.de -> Materials -> Lectures -> CogSci
NB: This chapter will only treat topics not mentioned explicitly in
    "Introduction to Neurobiology"

Some facts concerning Neurons:
 The human brain is made up of about 1E10 to 1E11 neurons. Each of
 them connects to about 100 to 1000 neurons. They take up 25% of the
 body's energy.

Lobes:
 The human brain consist of 4 lobes (Lappen):
 * frontal lobe (in front :-)
 * parietal lobe (in the upper back side)
 * occipital lobe (in the lower back side)
 * temporal lobe (in the middle)

Directions for descriptions of the body:
 * anterior -- towards the front
 * posterior -- towards the back
 * lateral -- towards the outside
 * medial -- towards the middle
 * superior -- towards the top
 * inferior -- downwards
 * dorsal -- towards the back (Ruecken)
 * ventral -- towards the belly

Contralaterality:
 Perceptions of the left side of the body are treated in the right
 side of the brain (and vice versa).

Hemispheric specialization:
 Both hemispheres of the brain are very similar and capable of nearly
 all functions needed. Nevertheless, each hemisphere specializes in
 different tasks (hemispheric specialisation, cerebral laterization).
 Usually the left hemisphere concentrates on language capabilities
 while the right side concentrates on spatial capabilities.

Word stem experiment:
 Task: The subject has to complete a word with several letters
 missing. The string can be completed to different words. Before doing
 the task, a story is read, where one of the possible words appears.
 Result: The subject will chose the word heard in the story.
 Conclusion: There exists implicit knowledge. Something has been
 stored in memory without explicit learning.

Representation of data in memory:
 There are two different theories on how memory stores data:
 * localized representation theory (each concept has its own memory
   area, if this area was destroyed, the concept could not be retrieved
   any more).
 * distributed representation theory (concepts are spread over the
   brain and cannot be given a locality)

Techniques for investigating cognition:
 * observing patients with brain lesions
 * imaging techniques (e.g. fMRI, PET)
 * other techniques such as e.g. single cell stimulation

Dissociation (Absonderung):
 It has been observed that a destruction of the medial temporal lobe
 leads to an anterograde amnesia (new data cannot be stored in LTM).
 In contrast, the destruction of the lateral frontal lobe results in
 a malfunction of the STM. Nevertheless, each lesion just influenced
 _one_ function of the brain.

H.M.-Phenomenon:
 H.M. was a patient whose medial temporal lobe had a lesion.
 Consequences:
 * STM was OK (1 minunte)
 * LTM for everything before 3 years before the surgery was OK
 * he suffered from anterograde amnesia (no STM->LTM)
 * nevertheless, implicit STM->LTM transfer (implicit learning)
   worked fine.

Brain map:
 It has been found out that different tasks stimulate different
 regions of the brain. Especially semantic and episodic retrieval seem
 to be different. This leads to the following model:
 * explicit memory (facts and events) involves the medial temporal lobe
 * implicit memory is
   * priming -- involving the neocortex
   * conditioning, i.e.
     * emotional responses (in the amygdala)
     * skeletal muscle responses (in the cerebellum)

Models for reading:
 * In Wernicke-Geschwind-Model, the visual information from the
   visual cortex (in the occipital lobe) is passed to different
   brain areas with different tasks (Wernicke's area, angular gyrus,
   Broca's area) and finally to the motor cortex in order to answer.
   Consequently, reading is represented in different cortical regions.
 * In the connectionist model, reading is represented by connected
   neurons firing on different stimuli. In the first layer, features
   of letters are recognized, in the second layer the letters
   themselves are recognized and in the third layer, words are built
   (bottom up). Now the word-neurons fire and inform the letter-
   neurons that a word has been recognized (top down). That's why
   the process is called "interactive".


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    EOF
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Allen Lesern wuensche ich viel Glueck bei den Klausuren und einen
guten Start in die wohlverdienten Semesterferien

           Fabian