Transcript Learning

Learning
• What is Learning?
– Relatively permanent change in behavior that
results from experience (behaviorist tradition)
– Can there be learning that does not result in a
change in behavior?
• Types of Learning
– Associative Learning (simple, passive, external)
– Cognitive Learning (complex, strategic, internal)
Associative Learning
• Classical Conditioning – associating two
stimuli
• Operant Conditioning – associating a
behavior and its consequences
Classical Conditioning
• Pavlov’s serendipitous discovery
• Associating 2 stimuli
– The first stimulus is “neutral” – does not
produce any response
– The second stimulus produces a reflex
(unconditioned) response
• After the 2 stimuli become associated, both
will produce the unconditioned response
Pavlovian Classical Conditioning
Before Conditioning
UCS
UCR
Neutral Stimulus
No Response
During Conditioning
CS
UCS
UCR
After Conditioning
CS
CR
Pavlovian Classical Conditioning
Before Conditioning
Food (UCS)
Salivation (UCR)
Tone (NS)
No Salivation
During Conditioning
Tone (CS)
Food (UCS)
Salivation (UCR)
After Conditioning
Tone (CS)
Salivation (CR)
Classical Conditioning to Cure Bed-Wetting
Before Conditioning
Alarm (UCS)
Wake up (UCR)
Full Bladder (NS)
No waking up
During Conditioning
Full B. (CS)
Alarm (UCS)
Wake up (UCR)
After Conditioning
Full Bladder (CS)
Wake up (CR)
Further Concepts that Apply to
Classical Conditioning
• Generalization: CR is given to stimuli that are
similar to the CS
• Discrimination: CR not given to stimuli that are
dissimilar to the CS
• Extinction: If the CS is presented repeatedly
without being followed by the UCS, the CR will
diminish or cease
• Spontaneous Recovery: Following extinction, the
CR will spontaneously re-appear after a delay
Classical Conditioning as Simple
Associative Learning
• Temporal Contiguity was thought to be
sufficient – the CS simply needs to occur
immediately prior to the UCS for
conditioning to take place
• Equipotentiality: any two stimuli could be
associated through conditioning
Equipotentiality Falsified
• Some stimuli are easier to associate than
others
• Taste Aversion – only foods become
associated with illness, not other stimuli
– Garcia & Koelling, 1966 – the “Sweet, bright,
noisy water study”
Garcia & Koelling, 1966
• CS = flavor, light, and click (sweet, bright, noisy water)
• UCS: 2 conditions
– Group 1: UCS = illness (from X-rays)
– Group 2: UCS = shock
• CR = avoidance (not drinking the water)
• After conditioning, tested which features of
the CS were associated with each UCS
Garcia & Koelling: Results
• Both Groups:
CS (sweet, bright, noisy)  CR (avoidance)
• Group 1(UCS = shock)
– Sweet water
 No avoidance
– Bright noisy water  Avoidance
• Group 2 (UCS = illness)
– Sweet water
 Avoidance
– Bright noisy water  No avoidance
Temporal Contiguity is Not Enough
• Contingency: The CS must reliably predict
the occurrence of the UCS (Rescorla, 1966)
• Informativeness: The CS must provide
new information for predicting the
occurrence of the UCS
Contingency (Rescorla, 1966)
• UCS = shock (S), UCR = fear
• CS = tone (T)
• Training: two conditions
– Random Condition: S TS S T TS S T TS
– Contingent Condition: TS
TS
TS
• Results: Rats learned to fear the tone only in
the contingent condition, when the tone
predicted the shock
Informativeness: Blocking
• If an organism has already learned that one
CS predicts the UCS, that will block the
conditioning of a new CS if the new CS
does not provide any additional information
• Example: Fear conditioning of a tone
blocks conditioning of a light
Blocking
Training 1
Training 2
Test
-none-
Tone & Light,
shock
(CR = fear)
Light 
Fear
Tone, shock
(CR = fear)
Tone & Light,
shock
(CR = fear)
Light 
No Fear
Rescorla-Wagner Model (1972)
• A mathematical model of the “strength of
association” produced in classical
conditioning
• Can account for all of the classical
conditioning phenomena we have just seen
• Uses just one single equation!
Rescorla-Wagner Model
ΔVn = c (Vmax – Vn)
V = the strength of association between a CS and a
US
ΔVn = the change in the strength of association
between the CS and US on a given trial
Vmax = the asymptote for CS-US association strength
after learning
c = rate of conditioning (how fast the association is
learned)
Associative Strength
between CS and US
Rescorla-Wagner Model
1.2
1
0.8
0.6
0.4
0.2
0
0
1
2
3
4
5
6
7
Conditioning Trials
8
9 10
Cognitive Interpretation of
Classical Conditioning
• Classical Conditioning is more than simple
association
• The concept of information could explain
contingency and blocking
• They are not just associating stimuli, they
are seeking information from one stimulus
to predict the occurrence of the other
Operant Conditioning
• The law of effect: behaviors that are
followed by good things happen more often
• Association: Things that occur together
become associated
Basics of Operant Conditioning
• Operant – freely emitted behavior operating
on the organism’s environment; NOT a reflex
response
• Reinforcement Contingencies – the
consequences that follow a behavior
– Reinforcement: increases the frequency of the
behavior
– Punishment: decreases frequency of behavior
Reinforcement & Punishment
•
•
•
•
Positive reinforcement
Negative reinforcement
Positive punishment
Negative punishment
Reinforcement Schedules
•
•
•
•
Continuous vs. Partial
Fixed vs. Variable
Interval vs. Ratio
Examples
–
–
–
–
Fixed ratio: vending machine
Variable ratio: slot machine
Fixed interval: checking mailbox
Variable interval: checking email
Explaining Complex Learning
with Operant Conditioning
• Secondary reinforcers - association
• Shaping – simple learning in small
increments
• Chaining – small increments plus
secondary reinforcement
• Language – association and reinforcement
(Skinner’s Verbal Behavior, 1957)
Learning that Could not be
Explained by Behaviorism
• Latent Learning – learning without
reinforcement (Tolman & Honzig, 1930)
• Observational Learning – learning without
behaving or being reinforced (Bandura, 1977)
• Overjustification – when rewards decrease the
frequency of behavior (but see Eisenberger & Cameron, 1996
for an opposing view)
• Language Acquisition – Chomsky’s critique
Latent Learning
Tolman & Honzig, 1930
Group 1: never a food reward
Group 2: always a food reward
Group 3: food reward after 10 days
10
9
8
7
6
Group 1
Group 2
5
Group 3
4
3
2
1
0
Day 1
Day 4
Day 8
Day 12
Day 16
Behaviorism Falls Short:
Language
• Chomsky: “Action in the past” as a property of
stimuli is sneaking mental representations in the
back door
• Association is insufficient to explain language
learning: The evidence points to learning RULES
• Evidence: Over-regularization (“goed”)
• Conclusion: Mere associations between words
can not explain language; any adequate theory of
meaning must hypothesize internal representations
of the rules of language (grammar)
So What was Behaviorism
Lacking?
• Symbolic Representation – we have
internal (mental) representations for things
in the external world
• Structure – we learn sets of rules for
combining symbols (e. g., grammar), not
just associations between pairs of symbols
Associative Learning Rises Again?
• LSA – Latent Semantic Analysis
– A theory of meaning, and a method for computer
analysis of the meanings of texts
– The meaning of a word = all of the words that co-occur
with it in a sample of written text (roughly)
– Meaning is just a function of associations of words, not
structure (syntax)
– How much of language meaning can LSA account for?
A surprisingly large amount.