Lecture 7 - Perceptual and Statistical Learning Flashcards Preview

COGS 101B Exam 1 > Lecture 7 - Perceptual and Statistical Learning > Flashcards

Flashcards in Lecture 7 - Perceptual and Statistical Learning Deck (30):

associative learning

both classical and operant conditioning (what goes with what?)


Both classical and operant conditioning require

the detection and prediction of contingencies (the presence of one thing relies on the presence of another) in the environment


For classical conditioning, the conditioned stimulus (CS) predicts the

unconditioned stimulus (US).


In operant conditioning, the operant behavioral response (R) predicts

the outcome (O).


In both cases, the association will become weaker (extinction) if

the prediction is less informative.


Paired-associate learning

is a testing procedure developed by Mary Whiton Calkins (1894).

• You repeatedly present pairs of items (usually
words or letter patterns) and allow the subject to build an ‘association’ between them.
- see how many repetions necessary to make the association

• Behaviorists later adopted a S --> R --> O model
in an attempt to explain this.

EX: trying to learn “rabbit” and “el conejo”

• Behaviorists say:
“rabbit” (S) --> “el conejo” (R) --> “good job!” (O)


paired-associate learning doesn’t quite fit the
behaviorist model (difficult to explain this kind of learning).

1) The learned association is bidirectional.
o Rabbit --> el conejo AND el conejo --> rabbit.
- you can trigger the other response even if it hasn't been reinforced

2) People use top-down information to facilitate the paired associations (Atkinson, 1975).
- people can use some kind of visual element (imagery) to combine and embody the two words: come up with an image that links the two
- problem for behaviorists because they argue it's only the enforcer

“rabbit” and “el conejo" (rabbit wearing an ice cream cone)


Classical and operant conditioning provided some great experimental techniques, but had serious theoretical flaws:

• Couldn’t readily explain individual or species differences in learning.

• Ignored top-down influences on learning.

• Contingencies can sometimes be bidirectional.

• In general, had difficulty dealing with computational complexity (hard to constrain learning outside of the laboratory).

However, there does seem to be something important
about how we TRACK REGULARITIES in the environment.


tracking regularities

we are good at seeing what goes with what


How accurate are we at detecting contingencies?

experiment by Smedslund (1963)

• nurses were asked to sort through 100 patient cards to see if a particular symptom predicted
the presence of a specific disease (outcome).

• Although the symptom is present more often overall, it has no real
predictive value for the presence of the disease (not contingent).

• 85% of the nurses determined that when the symptom was present,
the disease was more likely. This is an illusory correlation.

***shows that humans have a tendency to find causal links and if they're not there we impose them.***

we're constantly looking for association and we find them (whether they're real or not)


Much of our learning is driven by

detecting regularities in the
environment following repeated exposures/experiences (non-associative).
- the perceptual system just looks for regularities (id and distinguish things in the environment)


perceptual learning

we become better able to distinguish similar stimuli with experience.

• Chicken sexers are able to distinguish sexes at 1 day of age. Many can’t verbalize how this is done.

• This learning may be due to (1) mere exposure or (2) discrimination training.


In mere exposure learning (Gibson & Gibson, 1955)

there is no explicit training during the presentation of stimuli.
- learning to make visual discrimination just by seeing things a lot

• Participants were shown a series of cards and indicated when they
matched a target (circular scribble). No feedback was given.

• Participants were initially good at identifying the target, but mistakenly identified the similar items.

• With repeated exposures, they were able to distinguish the target
from very similar distracters.


With discrimination training

feedback is provided to
improve accuracy. This may combine perceptual learning with operant conditioning.


Fiorentini & Berardi (1981)

discrimination training

• trained participants to match complex visual images rotated by 1 degree.

• After initial training, they rotated the images 90 degrees…

By changing the orientation of the images (but not the
features of the image), performance dropped back to chance.

• This is an example of learning specificity.

• initial learning might interfere (when the stimuli are very similar)


learning specificity in perceptual learning

The training or ability does not easily generalize once the stimuli change in a
significant way.

you're learning to make just that one discrimination


perceptual priming

you are more likely to recognize (and
faster to process) a stimulus that has been seen previously.

• If you saw “Motor Supply?” on the first slide before lecture, you
would have been primed to use those words to complete the word stems.

• This can be done without conscious recall and is an example of latent learning. The priming had an effect, even though you may not have been aware of it and it didn’t have an immediate behavioral impact.

• This also implies a learning-performance distinction. Your internal system is changed as a result of experience (i.e.
you’ve learned something), but that doesn’t necessarily result in a difference in behavior (learning and behavior can be separated).


latent learning

priming had an effect, even though you may not have been aware of it and it didn’t have an immediate behavioral impact.

learning done without any sort of awareness and no conscious recall


learning-performance distinction

Your internal system is changed as a result of experience (i.e.
you’ve learned something), but that doesn’t necessarily result in a difference in behavior (learning and behavior can be separated).


behaviorists: learning is performance

but for priming studies

we can have latent learning (some change was made)


Nonhuman animals also show effects
of perceptual priming.

• Bond & Kamil (1999)

that blue jays were more accurate to detect a ‘moth’ if they had been primed with the image before.

• This kind of non-associative learning
seems ubiquitous.


the perceptual system is always looking for things

contingencies or patterns

even if we're not aware of it


statistical learning experiment on CogLab

In our visual experiment with statistical learning, you watched a continuous
stream of shapes. The stream contained four triplets of three shapes that
always occurred in the same order.
You were then tested with pairs of triplets and asked which one seemed
more familiar. There were three types of test stimuli.

• Always (A): Three shapes in the right order (e.g. ABC).

• Impossible (I): Three shapes that never appeared in order (e.g. ADG).

• Possible (P): Could appear by chance (e.g. BCG).


Our group data (n=173) indicates that you were sensitive to
the patterns:

• There will be individual variation, but the pattern is the same as the global data.

• You may not have been aware of the patterns. This would be an example of latent learning.

you didn't have to be actively making associations to produce contingencies through exposure


Adults are very good at extracting structure from a stream of stimuli.

Saffran, Aslin & Newport (1996)

asked if 8-month old
infants could learn word boundaries based purely on
sequential statistics.

• Infants listened to two-minute streams of four three-syllable
nonsense “words”.

• The only cue was that the transitional probabilities within words (da followed bi 100% of the time) was greater than between
words (pa followed ku 33% of the time).

Would infants notice the difference between words (bidaku) and non-words (dotigo)?

• Yes! Infants detected the difference between bidaku and dotigo (novel)

(as measured with preferential looking measures).

• This demonstrates that infants (at 8 months!) can segment words based purely on statistical cues, and with very little exposure (2minutes!).

• These are contingent probabilities, not simple frequencies.


Paired-associate learnin

(1) demonstrates that contingencies between stimuli can be bidirectional.

(2) They can also be affected by top-down processing (without explicit reinforcement).


Humans appear to be biased to see

illusory correlations where
none exist.

This means we track and learn regularities in the environment.


Perceptual learning allows us to make

fine distinction between

This can be done through mere exposure or
discrimination training.

However, it seems to be specific to the items learned and does not generalize.


Priming illustrates

how we are faster to process certain images if
we have had prior exposure

If there is a delay, this is a good example of latent learning and the learning-performance distinction.


Statistical learning (a kind of perceptual learning)

demonstrates how we can learn visual and auditory contingencies from the
environment based purely on probabilities.