4 - Configural Learning and Generalisation

Question 1

The Negative Patterning effect

Answer 1

a non-linear discrimination problem
Training:
A+
B+
AB-

Question 2

The Negative Patterning effect and the problem with the Rescorla-Wagner model of learning

Answer 2

The Negative Patterning effect cannot be modelled by the Rescorla-Wagner model of learning, because the R-W considers each stimulus independently, and cannot comprehend A and B together as behaving differently to how A and B would individually
Since A and B are both paired with + and -, it is impossible to learn, the model thinks AB must be twice as good as A or B
Running this in a simulation will show MORE outcome in the presence of AB, despite receiving the opposite training

Question 3

How to simulate the Negative Patterning Effect into the Rescorla-Wagner Model of learning

Answer 3

This problem can be circumnavigated by considering a common component of the two elements (X):
Training:
A+
B+
AxB+
- This works well when simulated

Question 4

Configural Theory

Answer 4

States that we learn separately about the cooccurrence of things, i.e. AB is learned as a separate entity X
- this allows us to solve a negative patterning problem

Question 5

Pierce’s configural model of learning

Answer 5

• predicts generalisation on the basis of similarity
• A is 50% similar to AB
• B is 50% similar to AB
• so the expected outcome of AB is 50% of the associative strength of A, and 50% of the associative strength of B
• so if the associative strength of A is 1 and B is 1, then the associative strength of AB is 1

Question 6

Model of generalisation on the basis of similarity

Answer 6

𝚫Ep = β (λ - (Ep + pSp’ * Ep’)

where:
(p = novel stimulus)

𝚫Ep = change in excitatory strength (outcome prediction)

β = salience of the outcome (importance)

λ = outcome that happens (outcome = 1, no outcome = 0)

Ep = excitatory strength

pSp’ = similarity to other stimuli

Ep’ = excitatory strength of other stimuli

Question 7

What is peak shift? How did Hanson observe this

Answer 7

Peak shift is the phenomenon when expectation for outcome is high for the trained stimulus, but highest at a stimulus slightly away from the trained stimulus, in the opposite direction to the discrimination stimulus

Hanson:

• he trained pigeons with different stimuli at different intensities of light
• S+ at 550nm
• S- at 560nm
  He found that following this training, pigeons will respond more strongly to light of 540 than S+, but less strongly to 530 and 520
• this shows a peak shift slightly past the positive outcome stimulus

Question 8

Peak shift compared to Rule and Generalisation

Answer 8

Mackintosh wrote up experiments by Aitken, showing that in humans, learning with natural stimuli with which they had experience (colour), there was no evidence of peak shift, instead evidence of Rule and Generalisation
- a rule had been learned (red means more food), and had been generalised to state that the redder it is, the more food (thus the peak of stimulation would be at the peak of redness, regardless of where they were trained to)

Question 9

Do humans show a peak shift effect?

Answer 9

Not in stimuli they are familiar with, i.e. colour
- instead, here they show Rule and Generalisation

But, humans were found to show a peak shift effect when learning with artificial stimuli

Question 10

Race Effect

Answer 10

• the tendency to recall faces of one’s own race more accurately than from other races

Question 11

Phonemes and their level of distinguishing

Answer 11

• the smallest units of sound that carry distinctions between meanings
• a 20ms difference across phoneme boundaries is easy to distinguish
• a 20ms difference within a phoneme category is hard to distinguish

Question 12

Perceptual Learning

Answer 12

Learning through sensory interaction with stimuli

Question 13

Conceptual Learning

Answer 13

Learning the broader principles in order to be able to apply them to other circumstances

Question 14

Can Perceptual learning facilitate discrimination between stimuli?

(Symonds and Hall’s flavour aversion experiment)

Answer 14

Yes
Shown by Symonds and Hall’s flavour aversion experiment

• two separate groups were given different training schedules for 4 days, where they would be given either AX or BX in the morning and afternoon, subsequently given AX with LiCl to induce nausea (AX+)
• there was also a control group that followed the same schedule, but during the training phase always had water (W)
• the test measured consumption of BX

AX = HCl and Saline (sour and salty)
BX = HCl and Sucrose (sour and sweet)

Intermixed (I) Group Training:
- AX / BX
- AX / BX
- AX / BX
- AX / BX
- AX+
Test:
- BX?

Blocked (B) Group Training:
- AX / AX
- AX / AX
- BX / BX
- BX / BX
- AX+
Test:
- BX?

Results:

• groups W and B both consumed similar amounts of BX
• group I consumed more BX

Because:
- group I had more opportunities (7) than group B (1) to determine the inhibitory association between A and B in relation to X (if one is present the other is not)

• thus group I developed a stronger inhibitory association between A and B
• whereas group B did not develop this inhibitory association, and could not compute the difference between AX and BX

Showing that our ability to discriminate between stimuli might be a learnt one
- not necessarily exposure to the stimuli, but learning the association between stimuli and outcome

Question 15

Rescorla-Wagner model’s interpretation of new stimuli Vs Configural Theory

Answer 15

R-W:
If trained for A+ and B+
the model predicts that AB? will produce ++

> associative strength of A to + is 1
associative strength of B to + is 1
so the associative strength of AB must be 2

Configural Theory:
- A is 50% similar to AB
- B is 50% similar to AB
> so the outcome expected from AB is 50% of the associative strength of A and 50% of the associative strength of B
> so in the same scenario as above, the associative strength of AB will be 1