Cognitive Science 207 Midterm Review session
11/18/04, Notes by Kate Lockwood

---

N.B. also look at the slides that were used, these will make much more sense in the context of discussion.

Big picture – main ideas covered in this class
 - Look at how computation can be used to capture cognition.
 - What can computation do?  What can it not do?
 - Idea of knowledge and representation being part of cognitive models.
- Markman representation – what it means, foundations
- Cyc representation model,
- Lenat – why representation is necessary to do common sense reasoning – large data base of knowledge

Qualitative physics
- a class of things that has a stronger organizing principle than Cyc – really need to organize facts by overlying principles.
- Qualitative process theory – everyday physical / natural world (not mental, not social)  
- there is a certain structure that can be conveyed by a small number of rules (ie causality)
- Want to capture in a common sense/intuitive fashion
- Natural language – key component of understanding the world of cognition

What is cognitive modeling?
Psychology – mind
AI – computer – how to make computers do what people do
Neruoscience – brain – brain is the mush inside your skull.  mind is your abilities / awareness.  neuroscience is concerned with how the brain itself works.  for example fMRI – look at where blood flows to brain in different tasks – how connections in brain work
Philosophy – everything – make theories about the world
Lingustics – language

have seen all of these things in the class so far

what are the GENERAL processes that compose these behaviors.  Run experiments on subjects build theories about how human minds work
Worry more about general processes than individual differences.  There will be variations, but they are from the same process (good memory/bad memory doesn’t matter, general recall process is the same)

understand cognition by creating computational models that operate on representations
this class worries about computational models and representations
want to convince you that this is a viable way to understand something about cognition

Minsky/Turing they want to prove that computation is reasonable as a way to understand cognition.  Not to say only way, but a very powerful way.  Computation is a very precise language.  If you can describe how something happens … need a very precise language so there is no confusion.  Come up with a very precise accurate description.  A program is just a very accurate description of a process.  Non-ambiguous.  You can get predictions out of it.  This is the key idea that cognition involves solving information processing problems (input – processing – output)  cognitive scientists believe in computation
Want to give examples of computational models.  
Computers do what we tell them.  This makes it hard to believe we can model minds.  We don’t do what we’re told to.  Much more complex in reasoning and behavior  But computation is more than that too

Don’t have computational models that are as powerful as anything we can do.
in very narrow domains maybe there is something impressive.
but computational models are doing cognition … they just aren’t powerful enough
making steps towards models of cognition
just want to capture some certain aspects of cognition
episodic memory is an important element of understanding language … this is a hypothesis … then you test it … it doesn’t have to always be as good as you or me at understanding language.

Analogy – structure mapping – psychological theory of how analogy works – 100s of experiments on how people do analogy – this model has held up to all of these results.  There is a model structure mapping engine – this is the program that implements the theory – made more projections based on how the program functions – take those back to the psychology library and gives more insights into how people do analogy

Representations – need to give some sort of facts, experiences, memories, etc.  How do you represent these for the computer.  Things are defined relative to other things.  Have circular definitions.  Yes, we don’t have everything, but you need to start somewhere.  Symbolic representations not sensory data.  Need symbolic description of things in the world.  

QP Theory.  About coming up with a language with which certain things can be easily represented – every day physical world stuff.  There is a certain structure that we can capture.  May be things that are missing, but fairly successful with a certain set of phenomenon.  Once you have the language it is very easy to describe new phenomenon.  Didn’t have to spend a lot of time thinking about what primitives to use, the tools / ontology was given to you.  A lot of work is already done

Reisbeck – encyclopedia vs journal.  Journal like experience in you life.  semantic memory vs episodic memory.  some people argue that we have both kinds of knowledge  maybe there are also other ways to represent knowledge – open mind project  - neural networks  Take some part of the world and create an artifact that some how represents it so you can do simulations in the representation and make predictions about the real world

Programs can reflect on progress and decide on courses of action, then it’s not doing exactly what you told it to – it is making decisions – you could be surprised by what happens.  Like HEURISTICS (the art of war).  Give general strategies to apply in certain situations.  The idea is to capture these general strategies and give them to the computer, this is a new way of thinking about programming

Representations
About organization – can’t just have a flat list of facts
Need a way to relate the things you want to talk about

Cyc

QUESTIONS
Vmodel – doesn’t need to look like the interface.  Define entities – define the parameters – parameters are constrained by processes – qualitative proportionality influence – not very complicated – not saying whether there will be on

Post a good model for both questions, but won’t be graded before the exam

How well do we need to know the CYC stuff?
know the difference between genls and isa
(genls dog mammal)

Collectison vs Individuals
Collections have instances (ie dog)
can be a part of another collection (ie dog and mammal) – all of these are collections
when you can have instances you have a collection
Individual is a single thing – do not have instances (ie Eiffel Tower – only one)
when ever you can say something is an example of something that is an isa relationship between the INDIVIDUAL you are talking about and the COLLECTION it is a member of
Instances have PARTS
Collecttions you don’t talk about Parts of.
collections capture structure, instances capture detail

Need to have clear distinctions and rules and structure because there are lots of people putting in knowledge so you need a standard form and so you can use the knowledge in programs

Can talk about dog legs, but this will be a collection.  Have a function that returns legs of dogs, so use function on an instance to get that’s dogs legs.
(forall ?d ?l  Universal quantification – variables that can stand for anything – when a variable is introduced it is automatically universally quantified, so we don’t need this part
(implies
(and (isa ?d dog)
(isa ?l (legFn Dog)) )
(hasParts ?d ?l))

?d and ?l are both individuals ?d stands for a specific instance of dog

dog, legFn are collections ?d ?l are individuals
with ISA first argument is always an individual and secong argument is always a collection
representation is painful

isa
X is an instance of collection Y
genls
Every instance of collection X is also an instance of collection Y
genls Dog Mammal
genls Tower FixedStructure
genls ModernMilitaryOrganization Organization

Y is a generalization of X
X is a specialization of Y

Collections of collections and collections of individuals
there are some collections whose instances are also collections
genls psychologist scientists
anything that is a collection like dog is an istance of a collection collection, but it isn’t an individual yet
when you want to enter something – is it an indivual or a collection
(isa psychologist collection)
collections usually the instance (like of dog) are specific individuals

( ______ CarEngine Car)
CarEngine is not a specific instance of car, so we know it is not isa
If we say genls then we are saying the CarEngines are a subset of cars, but this is not really true.  If we say genls dog mammnal then every dog is a mammal, but we don’t want to say that every car engine is a car
DisjointWith is two collections that have nothing in common like dog and cat.  Disjoint with is true because you can’t find an existence of CarEngine that is also an instance Car

QP Theory influence …etc … what is the name of this ontology … Process ontology Device ontology Device ontology – model a circuit individual components … battery ie.  
qualitiative proportionality – when one thing increases another thing does (moves in the same matter) also inverse qualitative proportionality (qprop-)
Derivative = flowrate influences the level of water in a
the flowrate is the rate of change in the level of the cylinder – this is called influences
qualitative proportionality

ARITY: number of arguments of a predicate
unary predicates takes one argument
binary predicates take two arguments ….
argIsa relationships constrain argument types.