The Early Years • 19thÎ20th century: all about the atom
Theories of Learning Wouldn’t it be ironic if I did a poor job teaching this lecture?
– Atoms associated to become molecules
• Ebbinghaus: paired nonsense syllables • William James speculated that Associationism associations in the world reflected in brain • Belief in a brain-mind correlation •Configurationism Learning = forming associations
The Early Years
The Early Years
• Competing view at the time • The mind/brain is an integrated system • Brain functioning – Specialized brain areas – Operate in concert (principle of “mass action”)
Associationism
Associationism
Configurationism
Configurationism
Before WWII: Gestalt • Koehler was one of the associationists • Recall his experiments with the chimpanzees on Tenerife – Learning to use tools wasn’t done by chaining Behaviourism S-R sequences – Involved seeing the problem as a whole
• We all know what the fate of Gestalt was Gestalt • I’ll mention Koehler again later
Before WWII: Behaviourism • Behaviourists did away with vague terms like association and replaced them with terms like S-R and Conditioned Reflex • Watson: All learning is S-R Behaviourism • The dominant theme for the 1st half of the 20th century Gestalt
1
Behaviourism: Clark Hull
Behaviourism: Clark Hull
• An attempt to extend behaviourist S-R paradigms to predict complex behaviour • Conditioning explains the mechanism of all learning • Wanted to learn what promotes learning S-R connections in the first place
• Drives (borrowed from Freud’s instincts) • Drives
– A rat won’t learn the maze if it’s never hungry
Behaviourism: Clark Hull
– Internal (hunger) – External (peer pressure)
• Reinforcement – reduces drive – Primary (food) – Secondary (money)
Behaviourism: Clark Hull • In the spirit of positivism, Hull proposed a mathematical model (easily testable) of behaviour
Response Threshold
– Hypothetico-Deductive Model
• Predict conditional probability of performing a behaviour: SER (reaction potential) • An example: P(BBQ some sausages)
Behaviourism: Clark Hull • What factors will influence SER? – Experience with BBQ food at reducing hunger? (SHR) – How hungry are you? (D) – How salient is the hunger? (V) – How much food are we talking here? (K) – Are we just tired? (IR) – Does our roommate usually give in and cook dinner? (SIR)
Behaviourism: Clark Hull • Let’s plug in some numbers to see a prediction: – Age 20: 17 summers of BBQ food (6 BBQs per summer), assume 50% satisfaction rate •
SHR
= 1 – 11-.03(17 x 6 x .50) = 0.97
– You’ve been to the gym: • About 80% as hungry as you’ve ever been (D=.8) • Your hunger is pretty salient (V = 1.0) • You’re really tired (IR = .27)
– Your moocher roommate does 1/3 of the cooking • Only 2 hotdogs left in the house (K=.8) • (SIR = .33)
.97 x 0.8 x 1.0 x .8 – 0.27 - .33 = .02
• Note that this formulation is probabalistic
2
Behaviourism: Spence
Behaviourism: Spence
• Recall from Hull’s theory: – SER (energy available to respond to stimulus) depends on both positive (excitation) and negative (inhibitory) values
• Gestalt phenomena had been difficult to explain using S→R paradigm • Spence showed how learned associations (habit strength, SHR) can explain these phenomena
Behaviourism: Spence S Training
Critical Test
M
• Train an animal to respond to a square of a certain size (16 x 16 cm) versus a smaller square (12.6 x 12.6 cm) • Over 100 trials: – Strong habit strength for larger square • SHR= 1 – 11-.03(100)= .999
– Strong inhibition for the smaller square
• What about the Gestalt problem?
Behaviourism Summary
S
M
L
1
.64
.08
1
.64
L
1
• Inhibition and habit strength generalize to similar objects as a function of similarity – Training: S is most similar to S, M most similar to M – Critical test: inhibition from S transfers to most similar object (M), habit strength for M applies to M, but also applies to similar object (L)
• These two behaviourists worked to expand the range of behaviours that their paradigm could explain • Make specific predictions using operationally defined inhibition and excitation
• Response to Gestalt objections to behaviourism
But then a funny thing happened… • A realization that there was one important behaviour that behaviourism couldn’t explain – Koehler argued (correctly) that association learning insufficient to account for language
• Even Pavlov came to see this as a problem – Second Signal System – stimuli were not raw physical inputs (phonemes, lexemes) but language embedded in verbal codes – A sort of linguistic configurationalism
Zombie Gestalt strikes again • Gestalt sensibilities continued to have an influence • Earliest behavioural experiments were ideally devoid of context – Ebbinghaus’ nonsense syllables
• One Gestalt legacy was to force experimenters to account for context – Hull acknowledged role of organism’s goals
3
Tolman
Tolman
• Behaviourism methods get a dose of context • Hull’s reinforcement driven S→R view contrasts with Tolman’s S→S view – Uses Gestalt principles to argue animals learn S-S connections without explicit biologically significant event to make learning occur
Tolman
– Knowledge of food locations not reinforced
• Hungry rats correctly navigated directly to food locations – If reinforcement required for learning, how did they learn the location when they aren’t hungry?
The Cognitive Revolution Cognitivism
You are here
– Learning is organizing things by their utility for achieving goals – trial and error is a looking back and forth to get the lay of the land in order to construct a solution – our cognitive maps are a record of goals and relevant paths
Tolman
• Learning without reinforcement • Satiated rats explored a maze
Behaviourism
• Cognitive maps • Learning is like mapmaking
• Cognitive maps contain expectancies made up of sign gestalts – Following landmarks – Changing landmarks disrupts navigation – Stores relative locations of objects/stimuli – Chain of associations – S→R alternative: learn associations between landmarks (S→S)
The Cognitive Revolution • Behaviourism’s failings most evident in problem of language acquisition • Had other failings too • S-R learning theory had nothing else to offer • Not a complete account of learning • By late 60s learning was being translated into the concepts of information processing • You’ll cover this at end of term
Gestalt
4
Connectionism – Hebbian Learning
Connectionism – Hebbian Learning
Connectionism 1990
You are here
• We talked about Hebb yesterday, so I went back and deleted this slide. • 45 second savings
Hebbian Learning • Based on correlational learning (James, 1890) – When two events co-occur or follow in succession, the connections between the neural representations of these events will be strengthened
• “Cells that fire together wire together” – Excitatory (+): 2 events occur together – Inhibitory (-): 2 events are mutually exclusive
Hebbian Learning • Phase sequences – Individual neurons participate in multiple cell assemblies • red associated with {cherries, fire hydrants, …} • Lets you minimize # neurons – Context constrains phase sequences to relevant associations
Hebbian Learning • Cell Assemblies – Interconnected cluster of neurons – The physiological mechanism for learning – Learning occurs at synapses between neurons • Information is not in the neurons themselves • Grandmother cell
– Strengthening of connections • additional connections or growth
Hebbian Learning • Reverberation – Cell assemblies may have reentrant connections – Facilitates single-trial learning • Associations a function of frequency
– Consolidation mechanism – Pattern completion – Mechanism for working memory • Computer memory works similarly
5
Hebbian Learning
From Hebb to Mastercard
A Simple Cell Assembly Demonstrating Reverberation
Hebbian learning at work • • • • • • • •
London, ON – Starbucks – 3.99 London, ON – Starbucks – 4.79 Cupertino, CA – Apple – 0.99 Cupertino, CA – Apple – 0.99 Cupertino, CA – Apple – 0.99 London, ON – Starbucks – 3.99 Cupertino, CA – Starbucks – 2.99 London, ON – Starbucks – 3.99
Mar 1 Mar 2 Mar 2 Mar 3 Mar 7 Mar 7 Mar 8 Mar 8
• Hebbian learning was first described almost 60 years ago • Mathematical modeling and computers incorporated these ideas into connectionism • connectionist networks can do simple human-like learning • Using Hebbian learning to detect credit card fraud: Priceless
Hebbian learning at work London
Cupertino Valid? Apple
Starbucks
Hebbian learning at work
Hebbian learning at work
London
London
Cupertino
Cupertino Valid? Valid
Valid? Valid
Apple
Apple
Starbucks
Starbucks
6
Hebbian learning at work
Constructivism
You are here
London
Cupertino Invalid Valid? Apple Constructivism 1990 Starbucks
Important Figures • John Dewey (1859 – 1952) – Philosopher @ Columbia University – Epistemologist – Philosophical founder of constructivism
• Piaget – Cognitive constructivist
• Vygotsky – Social constructivist
• Piaget & Vygotsky appear in later lectures
Constructivism • Constructivism: knowledge is a constructed entity made through the learning process • Knowledge can thus not be transmitted from one person to the other – it must be reconstructed by each person
• Knowledge is relativistic – It varies over time; it is something to be built up
Constructivism • Behaviorism: knowledge comprises passive, automatic responses to stimuli • Cognitivism: knowledge comprises abstract symbolic representations in the head of individuals – Computer metaphor (e.g., binary code)
• Both of these views see knowledge as an absolute entity – You either know something, or you don’t
Constructivism • People learn by leveraging existing knowledge • Thus, learning will depend on what else is already known – Best method for teaching will attempt to capitalize on existing knowledge – The reverse of the behaviourist approach – Echoes Gestalt ideas
7
The take home message Cognitivism
Associationism
Configurationism
Behaviourism
Gestalt
Connectionism
Constructivism
8