Cognition Final

Recency vs Primacy Effect

• Recency: small short term buffer (more recent terms)

• Primacy: rehearsal is required (first items said)

Atkinson Shiffron Model (interference)

• model has a large duration, going back and forth from STM to LTM

• Interference: competition between items preventing consolidation

  • Reconsolidation massed vs spaced training

• Encoding: consolidating a representation of info into memory

Forgetting

• Failure to Consolidate: no LTM

• Failure to Retrieve: in LTM

• Interference among cues

H.M. and types of amnesia

• hippocampus removed

• unable to form new memories

• LTM was unaffected and STM was unaffected

  • however he couldn’t consolidate (go between) because of lack of hippocampus

• Reterograde: LTM damage cant remember past

• Anterograde: consolidation damage cant make new memories

• Shows Procedural Memory is encoded seperately

  • HM could do learning task just like everyone else

Types of Memory (Declarative: episodic, semantic, procedural: mental imagery)

• Declarative (explicit)

  • episodic: first person memory of an experience

  • semantic/propositional: knowledge of facts

• Procedural (implicit)

  • how to do things - motor procedures

  • mental imagery: knowledge of appearance of things

Verbal Encoding, STM Capacity, Patient SF

• Capacity is 7 plus or minus 2

  • thats why phone numbers have 7 digits!

• Verbal Encoding: verbal distraction interferes with encoding while visual doesn’t

• Patient SF: taught to chunk rapidly

  • didn’t have more slots for chunks in memory just could fit more digits in each chunk

Imagistic vs Propositional Representations

• Imagistic: stores the sensory experience

  • visual representation in your mind: (image)

• Propositional: stores the abstract relation

  • “the cat was under the chair”

Mental Rotation

• mental analog of physical rotation

• spatially organized analog of a real picture is progressively transformed

Mental Scanning

• Mental distance is an analog of real distance

• more map distance = longer it takes

• Proof mental imagery exists

Mental Imagery Debate

• Kosslyn

  • pro mental imagery

  • evaluation → visual system

• Pylyshyn (Rutgers)

  • representations of images are propositional/descriptive

  • scanning results due to cognitive expectations of subjects

Property Inheritance

• If x is a y and y has property z then x has property z

• direct property wins!

• inference ability = efficient form of reconstructive memory

• Hierarchical Structure (top is superordinate to bottom subordinate)

Induction vs Deduction

• Induction: specific to general

  • if it rains on saturday and it rains on sunday it rains on weekends

  • a guess/generalization

• Deduction: general to specific

  • all men are mortal, socrates is a man, socrates is mortal

  • logical conclusion

David Hume

• will the sun rise tomorrow? How do you know?

• Hume: belief that the sun will rise tomorrow isn’t logically sound and only whats likely

• Induction not deduction: not guaranteed to be valid

Classical vs Modern View of Categories

• Classical View

  • categories have logical constructs and definitions

  • definitions have necessary and jointly sufficient features

  • clear cut categories but is a hotdog a sandwich?

• Fuzziness and Family Resemblance

  • most mental categories have fuzzy boundaries with family groups

  • Ex. game, furniture

• Exemplar Model

  • comparing information to stored examples/exemplars in brain

  • no generalization/abstraction

Classical Concept Learning Experiment

• Find the rules!

• Ex. Squares or small and green shapes

• Conjunction was easier than disjunction

• assumed concepts are defined with logic

Prototypes

• Posner and Keele (1998) - prototypification = “normal form”

• Ex. prototype of bird

  • each feature in proportion to its prevalence; not necessary/sufficient

• distinguishing core meaning and identification purposes

Geometric Model + Multidimensional Scaling

• Geometric Model

  • analogous to proximity in some mental space

    • dissimilarity is more distance, similarity is closeness

• Multidimensional Scaling

  • ranking dissimilarity among set of items finds positions in imaginary space

    • x axis is money y axis is sporty

Euclidean distance vs city block distance

• Euclidean (as the crow flies) - purple

• City Block Distance - pink

Distance Axioms

• Symmetry - distance from A to B = distance from B to A

• Triangle Inequality - distance from A to B + distance from B to C has to be greater than the distance from A to C

  • shortest distance between 2 points is a straight line

Similarity and Distance Axioms

• Similarity DOES NOT obey distance axioms

  • geometric model is wrong

• apple compared to pomegranate was more liked compared to pomegranate vs apple

• also failed triangle of inequality

Contrast Model

• Similarity of A + B = Features shared between A and B - Features of just A - Features of just B

  • uses featural model to contrast features

  • can violate triangle of inequality but is still used

Connectionism/Pandemonium

• Connectionism

  • artificial neural network is the alternative approach to cognitive architecture

  • rooted in empiricism/associationism

  • ties to neuroscience

• Pandemonium

  • lots of small individual parts and layers make up larger system

McClelland and Rumelhart

• parallel distributed processing book

• dumb device that recognizes its stroke

• Flaw: cant possibly be infinite neurons to make infinite sentences

Training a Neural Network

• drawing examples from a training set

• Supervised: know what the output should be by comparing to an oracle

• modify parameters to improve performance by altering connection weights

• repeat many times and reevaluate

• approximates target function

Backpropagation

• feed input and get output

• oracle compares actual output to target output

• compares discrepancy (error) between target/actual

• Reduce error and repeat

Principles of Connectionism

• biological plausibility - should be based of the brain

• parallel computation - brain = many units working at the same time

• distributed knowledge - all info is on the weights (spread)

• 1 universal learning model - 1 mechanism for all parts

Symbol System vs Connectionism

• Symbol System - rationalist/nativist

  • brain uses rules with symbols

  • different learning mechanism

  • some knowledge is innate

• Connectionism - empiricist/associationism

  • rules/symbols = epiphenominon

  • 1 general mechanism

  • all knowledge is learned

Long Term Potentiation

• how one synaptic connection excites another

• Excitatory: raising excitatory weight

• Inhibitory: raising inhibitory weight

Deep Learning

• more layers with a huge database for training

• Is it like human learning?

  • yes - induction/neural network

  • no - need billions of examples and is bad at generalizing

Reasoning (prescriptive vs descriptive)

• going from premises (existing beliefs) to conclusions

• Prescriptive - how should reasoning work?

• Descriptive - how does reasoning work

Induction

• “squirrels like nuts, squirrels are like badgers, badgers like nuts”

• plausible not objectively valid

• Thomas Bayes: we can quantify degree of belief as probability “common sense calculations

Deduction (Modes of Deduction)

• Modus Ponens: If A then B, if A is true, B is true

• Modus Tollens: If A then B, B is false, A is false

• A → B (A is antecedent and B is consequent)

• Probable Modus Ponens: P(AIB) is high = P(B) is high

• Probable Modus Tollens: P(AIB) is low = P(B) is low

Wason Selection Task

• Vowel on one side # on the other

• Cards say: U K 4 7

  • most people want to flip U and 4

  • Logically you should flip U and 7

• If under 18 you have to drink coke

  • most people want to check under 18 and wine drinkers

  • logically you should flip under 18 and wine drinkers

• Conclusion: humans are not very good at reasoning, deductions aren’t always useful, many inferences we make are not deduction

Probabilities

• P(A) = probability that A is true (true is 1, false is 0, P(A) = between 0-1)

• P(-A) = 1 - P(A) probability of not A is the probability of 1-P(A)

• P(A and B) - probability that A and B are true

  • Conjunction Rule: probability of A and B has to be greater than 0 and less that P(A)

• P(A V B) - probability of A or B

  • Disjunction Rule: probability of A or B is greater than the P(A) but less than 1

Conditional Probability / Reasoning

• P(AIB) = probability of A given B is true

  • P(AIB) = probability of A and B / Probability of B

  • P(AIB) x P(B) = P(A and B)

Bayesian Inference

• allows you to compute how strongly to believe H as a function

• form beliefs in an optimal way, normative reasoning, objectively correct

• Likelihood - degree which H fits evidence

• Prior Probability - how likely H before evidence

• Posterior Probability - probability of H given D

Are People Bayesian? (Conjunction Fallacy)

• Economics - people make optimal use of information acting in own rational self interest → Bayesian!

• Conjunction Fallacy: more people are likely to say Linda is a feminist and bank teller rather than just a bank teller

  • subset should be smaller than the whole thought

• Perhaps perception is but cognition isn’t but its still highly debated

Representativeness Heuristic

• Heuristic: approximate strategy for problem solving thats easier than optimal strategy

• In this scenario subjects typically say 80% because people ignore the prior

• people dont use bayes rule but instead use heuristics and biases

• Prescriptivists: this kind of reasoning is genuinely defective

• Adaptive Strategies: heuristics work well in real ilfe

Fallacies

• Law of small numbers - people use small representations to draw conclusions

• Hot Hand fallacy - one side of the coin is hot so p(H) is more than 50%

• Gamblers Fallacy - HHHHHTHHHH tails is due so P(T) is more than 50%

Expected Value

• long run average value (EV=sum of xp(x))

• 50% chance of one inch of rain

  • 0.5×1+0.5×0 =0.5 inches

• The rational price of a bet is its expected value

Subjective Utility

• how desirable something is to someone

• Maximizing expected utility is how people make decisions

Risks and Risk Aversion

• Do you pick 100% chance of $10 or 50% chance of $20

  • expected values are the same

• Risk Aversion: preference for certain gain over chance of loss

• 50% chance of $20 or 99% chance of $12

  • EV 1 = $10 EV 2 = $10.88

Temporal Discounting

• Do you want 90% chance of $11 now or 90% chance of $15 in a year?

• Most people choose the first option because subjective devaluation of future value

Preference Reversals

• if you prefer A to B it shouldn’t matter the other choices

• You have a choice between chocolate, german chocolate, and vanilla, you prefer chocolate but you are scared because of similar option so you go with vanilla

• Allais Paradox - offered 50% of $100 or 100% of $50 people choose 2 even though both EVs are $50;

  • however they want 5% of $100 instead of 10% of $50 even though they are the same too

Framing Effects

• changing the wording of a question to alter decisions

• You have 600 people on a sinking ship do you

  • save 200 or kill 400? Same thing hahaha

Sunk Cost Fallacy

• continuing to invest in something due to the resources you have previously spent on it

• past should not impact your choice but the framing of the past does

Bounded Rationality

• humans are rational but limited in terms of info, time, comp resources

• Satisficing: choosing the first acceptable pick or deciding based on one issue only

Chomsky vs Skinner

• Skinner

  • humans learn language via reinforcement

• Chomsky

  • infinite compositionality and productivity proves that Skinner was wrong

  • not all sentences have been said… so how could they be reinforced?

  • “colorless green ideas sleep furiously” makes no sense but is still valid

  • all leads to an abstract rule system which makes up grammar

Ape Language

• Skinner - if correct apes should be able to learn ASL

• Chomsky - if right they shouldn’t be able to because they need innate structures

• Result: apes can learn vocab but they learned little syntax

Competence vs Performance

• Competence: abstract knowledge held by a competent speaker, what you implicitly know about distinguishing between languages

• Performance: how this works in practice, actual mechanisms for carrying out speech/hearing

Prescriptivist vs Descriptivist

• Prescriptivist: rules tell you how to do language correctly “don’t split infinitive”

• Descriptivist: rules tell you how its actually done

Universal Grammar

• all natural languages have common structures which are innate (Chomsky)

• all have words and sentences

• all have nouns, verbs, and other parts of speech

• systematic variations of each other (SVO, SOV, OSV)

• how to form a question (insert SV, Auxiliary verbs)

• which words modify the others (the boy with the dog was happy)

Is Language Innate?

• humans have unique ability to learn language without overt instructions

• some aspects of language have a critical period

• Poverty of the Stimulus - linguistic input isnt enough for a child to learn rules of a language

  • Im going to eat lunch → im gonna eat lunch

  • Im going to New York → Im gonna New York

• Only universal grammar is innate while specific details have to be learned

Subjects in Linguistics

• Phonology - study of sounds (consonants, vowels, variations)

• Morphology - construction of words out of units that carry meaning (morphemes)

• Syntax - ordering words to form sentences (grammar)

• Semantics - meaning and logical form

• Pragmatics - practical aspects of conversation (what did someone mean? what do i say next?)

Syntax

• abstract rules for legal sentences that sound normal and make sense to a native speaker

• generative grammar - produces all and only the legal structures in the given language

  • ex. rewrite rules: The dog chased the cat. The tree chased the cat. The noun chased the cat

  • Recursion adds an extra adjective every time you do it

• Tone localization experiments confirm “psychological realities” of boundaries

Syntactic Ambiguity/Garden Path Sentences

• sentences can be parsed i 2 or more ways

  • ex. “the girl hit the dog with the stick”

• Garden Path Sentence - principle of minimal attachment

  • ex. “the cotton shirts are made from comes from Alabama.”

Phonology

• Sounds of a language; individual sounds = phonemes

  • “Tooth” = /t/oo/th

  • English has 44 phonemes

• Phonological Parameters

  • Manner of Articulation - stop (p,b,t,d) or fricative (f,s,th)

  • Place of Articulation - bilabial (p,b) or labiodental (f,v)

  • Voicing characteristics - voiced/voiceless (f/v, s/z, th/th)

Categorical Perception

• elevated sensitivity to sensory differences across categorical boundaries

• In continuous categories (color spectrum) people see categories instead of one fluid scale

• after critical period - can hear distinction between categories but not within

Semantics

• lexical semantics - accessing mental lexicons

  • ex. man gave woman a hand → knowing meaning

• Reference

  • the man ate the sandwich he made (he and the man corefer)

  • the man looked at him in the mirror (the man does not refer to him)

• Locigcal Form

  • the boss takes her coffee with sugar

    • she usually takes it with sugar OR she is currently drinking coffee with sugar

  • Everyone has a cell phone

    • all people share 1 phone OR all people have their own phone

Phrase Semantics

• brick house/rabbit house

  • house made of bricks vs a house where rabbits live

• Implicature: inferred propositions that are understood as premises

  • Where in England did he visit implied that he visited England

  • Do you beat your wife on Sundays implied that you beat your wife other days

Morphology

• pieces of words that carry meaning (tense, gender, number)

• roots, prefixed, suffixes

• Cat+s = 2 morphemes

• Potato = one morpheme

• re+paint+s = 3 morphemes

Past Tense War

• Morphological Rule - add -ed

  • talk → talked (+1 sound)

  • aid → aided (+1 syllable)

  • bug → bugged (d sound)

• Lots of exceptions: is/was go/went read/read but not need/needed

U shaped Curve and Over-regularization

• Overregularization: child learns the rules and over-applies it as the learn the exceptions

  • this is why there seems to be a dip in childs progress of language learning

• Wug Test: today I wug, yesterday I _____

  • any common speaker can pass this test

Pinker + Price’s Rebuttal

• Network generalizes incorrectly

• some make mistakes no kid does (squat → squakt; mail → membled)

• cant handle homophonous verbs (fly/flew; but flied out)

• Mis ordering issue and rigged U shape curve (regulars first - irregulars later when its actually a mix of both)

Past Tense Neural Network

• input/output by backpropagation on surface forms

• no rules, no morphemes, no distinction (1 mechanism)

• application of a general mechanism

• replicated U shaped learning curve