PSYCH 354- in class notes

Titchner

school of psych

• unified psychology through its methodology

• invented new techniques to turn experiments to uncover properties of mind

what was the crisis in psychology

no one knew what psychology was, there was too many schools of thought

• it made up of different perspectives and ideas rather than a singe concept

how did psyhology get furhter fragmented

people emphasized content therefoe deemphasized the foundational concepts that tied psychology together

• heavy focus on isolated facts miss out on foundational elements that unite the discipline

difference between cognitive science and psychology

cognitive science is more interdiciplinary and requires research from all discipines involved, whereas psychology is more focused on studying human behaviour and the metnl processes behind those behvaiours

information processing

a feedback loop originating in cybernetics

• symbolic processing within a computer

cybernetics

the idea that an agent acts in the world there is a consequence that acts back on that agent from the world

how is information processing used in cognitive science

it is used as an abstract notion of information processing that glues togetehr the disciplines

• acts as a common language allowing them to communcate across disciplines

• by using the analogy that the mind is a computer it turns the mind into a inaminate abstract object into a physical componet an mkaes it easier to talk about

the information processing hypothesis

the human mind is a complex system that receives, stores, retrieves, transforms, and transmits information

• or cognition is computation

• or cognition is information processing

how deos the brain take on connectionist foundations

the brain process information but there are no rules or symbols but instead symbolic interactions within the brain

• processors send signals to other processors and so on…

essentially what is the main idea of embodied cognitive science

by acting on the world, manipulating the world

• the world ofeeres potential actions u can take part in (ka affordances)

Man as Machine → Rene Descartes

believed there was competition between physical vs non physical account of the mind- that are interrelated

• moved towards mehcnaizing thought

logicism → george boole

philosphy and a mathematical perspective into what human thought might be like if explained mechanically

• introduced mathematical rules for combining and evaluating symbols

• human thinking is the same as logical operations defined in his algebra

20th century truth tables

evolved boolean logic

• logical operations over 2 symbols, that can either be true or false → assigns truth values, while lgogical expressions assign truth value based on telationship between variables

• 2 varaibles alwyas has 4 possible states

what was Claude Elwood Shannon’s contribution

used boolean algebra to test the circuit before building it→ more efficient and cost effecitve

• translated mathematical logic into physical, electrical switches

• write circutis based on boolean algebta to create physical circuits→ output is the sam even if physical input is diffrent

2 levles of description

physically or describe the funciton its computing

algorithmic level

the basic components- the order of the sequence can change the outcome of the algorithim

• doesnt ask what or why, but rather answer the how question → what stteps are carried out in what order to permit the output of this combinationn lock

architecture

the univeral builing blocks, required to bring sequence of steps in algorithim to life

• symbolic account

• primitive information processing components- built into computer

functional analysis within the architecture

what each component does (funciton) → physical nature ignored

fucnitonal analysis

define a big task that an information process is carrying out → collecting experimental evidence to create a sequence of more basic/simpler functions ta if carried out in order bring the big function to life

• tote umit→ pure info processing, purely funcitonal

• test → treated as information, being either true or false

ryles regress

functional not physical

• keep breaking down funcitons into simpler parts, decompostion further and further in a never ending loop

problem with ryles regress

if theories only use functional terms u are unable to explain anything

• explaining something leads into making more functions

how did robert cummins explain how to escaepe ryles regress

introduced the the idea of subsumption

• after going through steps of ryles regress (definition, analysis) there comes a point where breaking down any further is redundant and must use subsumption

causal subsumption

building a simple machine to carry out a simple function (ex. neuron- either fires or it doesnt- simple)

• brings our simple fucntions to life, and fucnitonal description becomes a dunctional explnation

multiple realization

different physical things can bring the same funciton to life

• shows that physical properties dont matter as long as theyre on the same architectural level

• ex. a relay, vacuum tube, and transistor all perform the same funciton but all look differently

what are the multiple levels of investigation

• computational

• algorithmic

• fucntional architecture

• implementational

*levels ask diff questions, methodologies, angague, and diff types of training. (one discipline is more suited for certain levels)

computational level

define, typically with math

• asks what infromation processing problme is being solved

• answered with formal proofs

algorithmic level

brings fucntion to life

• asks the question of what sequence of steps- what algorithim- is being used to solve the information processing problem

• answered with experimentla studies of behaviour

• more than 1 algoirithim can perform at 1 computation

  • specifying an algoithim or program - infer the underlying algorithim

architectural level

asks what primitive operations is the systems algorihtim constructed→ decomposed algorithim into basic primitive info wired in the device

• answering this question involves a combinaiton of bevaioural observations and the methods of cognitive neuroscience

• what wired inprocesses are avaialbel to carrry out the algorithim

implementation levle

asks what physical porperties are responsible for birnging the archtectures primitives into being?

• answeing this question appeals to the narueal sciences

the explanatory goal

computational → algorithmic → architectura → implementational

• goal is to finf the 1 architecture for the 1 participant carrying out the algorihtim carrying out the computation we are studying

• connectivism approach

classical approach

rule governed manipulation of symbols

disembodied mind

claim that the mind is completely distinct from physical world

• words like mind, soul, spirit are synonymous in thhat theyre completley separate from physical world

mechanizing the infinite

the menral is infinite while the physical is not

• this difference cuases the possible seperation between the 2

the turing machine

device was central to making digital computers

• question answering machine

• physical device that does the rule governed manpulation of symbols

implication of the universal machine

can do anything that can be computed

• any algorithim perceived a turing machine is able to carry it out

physical symbol systems

the idea that finite physical device that can generate an infinite variety of behaviour

• brings to life idea of information processing

• rules in machine head

• device can be used to explain human cognition → brain as a physical symbol system that does cognition, perception, etc.

recursion

concept from math that states how finite things are able to produce infinite amounts of behavior

context-free grammar

rewrites rules

• can generate inifinite choices since its recursive

comparitive cognitive science

the comparison between theory/model with the participant

Eliza and the turing test

demonstrated weakness of turing test

• since it was able to pass even though it was designed specifically to not understand language

weak equivalence

2 systems are weakly equivalent if they produce the exact same input-output mapping

• they solve the same overall information-processing problem or compute the same function

• established strictly at the computational level

strong equivalence

when 2 systems are strongly equivalent if they compute the same input-output funcition and do so using the same algorithim grounded in the same architecture

• to prove this researchers would collect specific behavioural artifacts and avidence

• requiewa identity across multiple levels of investigation→ computational, algorithmic, and architectural

why do they run experiments in cog sci

to establish strong equivalence

first order

refers to the primary input-output mapping or direct stimulusresponse behvaiour of a system → focusing strictly on the final result generated when given a specific input

second order

refers to the unintended behavioural side effects or byproducts of processing that arise from the interaction between internal data structures and the active rules manipulating them (processes

• ex. time, difficulty, states, errors

descartes

nativism, innate traits

• new knowledge gained by rules of logic

• cartesian

locke

empiricist, connectionist

• wanted to get rid of nativism

why cannot logical theories explain human cognition

since the brain doesnt always use logic

how does procesisng units learn form mistakes

find the difference between the desired/teaching value and the actual output produced to find the error

• use the error to train the wighted connections (bigger error= bigger weight change)

• learning descreases error and makes them more accurate

pattern recognition

common in networks trained by cognitive scientists

• processers as feature demons → if it sees its feature it shouts (aka activity)

physical symbol system hypothesis

states that a physical ysmbol system has the neccessary and sufficinet means for general intelligent action

• human brain does informatino processing diff froma computer, therefore computer cannot be used to model (slow- one rule at a time, and brittle- one lesion can have massive effects)

the PDP alternative

since the brain operates in parallel, better to use parallel distributed processing as a model

• knowledge distributed through maze of connections, each connection does not have te same weight and therefore can work even if one area is damaged

• argued to be able to fix classical problems

perceptron mark I

first learning machine

• empirical device→ starts as a blank slate that learns

perceptron structure

12 input units, 1 output (no hidden units)

• old connectionism→ stimulus-response machine (learns due to its simplicity)

perceptron limitation

book states that theres lots of problems the perccptron cannot solve but people can

• therefore using perceptron as a model for human cognition is not accurate and should be abandoned

linearly separable problem

able to seperate true and false with one line

• if linearly-separable perceptron is able to solve

• ex. AND statements

linearly non-separable problems

unable to be lineraly separated, need more than 1 cut

• perceptron unable to learn how to solve this problem due to absence of hidden units

• ex. exclusive OR or XOR statements

perceptron legacy

showed that u can train simple neural networks

• to solve linearly non-separable problems u would need at least 1 later of hidden units

modern networks

powerful empiricist networks

• connectionist revolution, artificial neural networks come back

hidden units

able to find relationships between first order properties

• 2nd order properties are more complex and allows them to solve more complex problems

• more of them = able to make more cuts in the space

  • are able to change the shae of the pattern space so now one slice separates the white patterns

credit assignment problem

when this network makes an error, how much is the error due to each hidden unit

• solved this issue using backprop

how to backprop

computational analysis done with step function--> discontinuous

• replaced with continuous approximation to be able to calculate derivative to calculate the slope

• translating calculus into an algorithm-- hidden unit error is a signal sent backwards to the hidden units from the output units

classical sandwhich

sense→ think → act

• no direction connection between sense and act

• sense = input

• think = information processing (hidden units) → manipulatee what we have sensed through processing

• act = output → acting on the world

abdandoning classical sandwhich

sense → act

• get rid of thinking and planning - sense and act have direct connection

• what u sense (perception) directly impacts action ( behvaiour)

what are the 3 themes of embodied cognitive science

• conceptualization

• replacement

• cosntitution

conceptualization

ur understanding of the world depends on nature of ur body-- everyone has a different view of the world (concepts of the world depend on the body)

replacement

take cognitive resources (ex. memory) and replacing it wtih external resources (ex. notebooks) that are present in the world

- replace internal cognition with externl cognition

constitution

the world is more than a source of stimuli

- the world is part of ur mind

feedback

actions are diriven by different sense-act cycles

• change show u act but also can change how others agents act

• change the world, change the behaviour

affordances

potential action done by an agent that is offered to the agent by the environment

• shape of body deterins the actions u can perform on the env and vice versa

• interaction between sensing and acting → different agents will think in different ways since body is interacting in the world will result in different affordance than another agent

cognitive scaffolding

thinking in embodied cog sci happens outside the skull

• thinking by maniulating the world

• use the world to replace some internal processing to relieve cognitive demans ex. index cards

stigmergy

working together on mutually beneficial task, without communicating, world communicates to the agenrs

• sense act refelxes

• no thinking involved

• finite set of reflexes

reverse engineering

look for second order effect to do a functional analysis

• take it apart to see how it works

forward engineering

abstract level → architecture

• build model first then collect data

verum factum principle

invented philosphy of law

• only way to understand things is to build them

• intorduced us to forward engineering