Lecture 11. Learning and explanation Flashcards
Does Learning involve attention? Yes! But to achieve our goals, what must we do?
Overcome initial fascination (attention) to salient but irrelevant attributes
What was the strong elements of the waggle dance?
It only contained the necessary information: direction and distance to food
There are four experimental effects which show us that attention is important for learning:
- what must we trade off between?
- when all of our attention is used for one cue so another is ignored
- when cue predicts an unusual event so it becomes more dominant
- appreciating different attentional rules
- validity and salience -
Salience is how much the cue grabs our attention all other things being equal. In the absence of validity high salience cues will attract attention - Blocking
- Highlighting
4, different complexity in attention - changes in understanding of learning theory
What are the rules for validity vs salience in regards to utilisation?
Utilisation is like validity but concerns how much people will learn about or use a cue on which to base their experience
We know that validity competes with salience right?
- Increased validity leads to _____ utilisation.
- decreased validity leads to _____ utilisation
- Increased salience leads to _____ utilisation
- Decreased salience leads to _____ utilisation
- Validity and salience ____
- Increased utilisation of one cue _____ utilisation for other cues
- Increased
- Decreased
- Increased
- Decreased
- Validity and salience interact
- Decreases
What type of attention is this? Early: A -> X Late: A.B -> X, C.D -> Y Testing B.D -> Y Y is preferred, B is blocked, attention goes to what is cued by D
Attention is shifted to A because it is important for predicting X
No Attention is left for B when A.B are paired
Cue D drives the final response
This is blocking
A -> X
A.B -> X
C.D -> Y
B.D -> Y
Blocking
A.B -> X
A.B -> X
A.D -> Y
B.D -> Y
Highlighting
What is simple learning theory?
The process of higher co-occurring cues-outcome combinations leading to strengthening of learning
So, simple learning theory is that co-occurences lead to strengthening between cues and outcomes, what does additional learning theory add to this?
The ability to differentially weight cues according to their relevance
What did the unidimensional rules model by Kruschke show us?
Hint: it was to do with attentional learning theory.
That attentional learning theory accounted for the fact that participants were slower to determine dots that involved two dimensions than dots that involved one dimension.
While the occurrence remained the same, we were able weight the filtration condition more relevant
What is a learning mechanism that is not accounted for by simple learning theory or attentional learning theory?
One shot specials!
One shot generalisation means we can get a good general idea about an object and its associations by seeing it once.
What does one shot generalisation utilise?
That expectations provide hypotheses which we use to:
- explain that we’re trying to learn
- evaluate our experiences (and therefore update our expectations)
Through this, we learn by exclusion and we base it on what we already know and expect
Sometimes, designs do not benefit from one shot design or from generalisation. What people think they know is often more than what they actually know. What is an example of this that was used in the slides?
The complexity of a helicopter rotor system
okay, we’re looking at blickets, which involves a box where when you put stuff onto it, and if that thing plays a sound, then its a blicket . What does that have to do with?
Learning and prior knowledge!
- What is a one cause condition?
- What is a two cause condition?
- What is backwards blocking?
- What is indirect screening off?
- Object A activates the machine by itself. Object B does not activate the detector by itself, Both objects activate the detector. Children are asked whether each object is a blicket. Children only label the first object as a blicket
- Object C activates the detector by itself (demonstrated 3 times). Object D does not activate the detector by itself the first time. The second and third time Object D activates the detector by itself. Children are asked whether each is a blicket. children label both objects as blickets
- Both objects A and B activate the detector (demonstrated twice). Then object A activates the detector by itself. Children are asked if each is a blicket. Children then only label object A as a blicket
- Both objects c and D activte the detector (demonstrated twice). Object C does no activate the detector by itself. Children are asked if each is a blicket. Children rate object D as a blicket
Expectations! What are they?
What can they be considered in terms of how we see the world?
Expectations can be considered to be hypotheses that we use to evaluate information in the world
In summary in regards to learning.
Learning depends on prior ___1.___.
A set of hypotheses make predictions for the types of ___2.___ you are likely to ___3.___
Upon observing the predicted events, those hypotheses are ___4.___; other hypotheses are ___5.___
1. 2. 3. 4. 5.
There are two ways prior knowledge influences our hypothesis. Here are the descriptions of these influences -
Only hypotheses which explain the data are plausible candidates for the an explanation. “Once you eliminate the impossible, whatever remains, however improbable, must be the truth.”
If one hypothesis clearly explains the data, then other candidates are considered less likely. “If one suspect confesses, then we let the other suspect go”
What are the names of these?
Holmesian Deduction
Judicial Exoneration
while holmesian deduction and judicial exoneration influence our hypotheses, what is the best method of arriving at a hypothesis?
What should we note about this and other other hypotheses?
Inference!
“I have some prior beliefs in a hypothesis H which predcits that I should observe some data D.”
“If I observe D, then H is supported”
Other hypotheses which also predict D are supported but “i dont believe them as much if my prior beliefs in the alternative hypotheses are smaller than my beliefs in H”
What was the deal with the peptic ulser?
How was it disproven that it
was caused by stress?
What was the deal with bacteria?
proving an alternate hypothesis!
it was proven that bacteria could live in the Stomach so marshall injested the bacteria, contracted an ulcer and then treated it with anibiotics
Baye’s Rule!
What is Baye’s rule?
It is: Your updated beliefs should be proportional to your prior belief times the likelihood of the observed (i.e. how much did your hypothesis actually predict the event?)
So, how much should I update my belief in a hypothesis after observing some data?
Updated belief = Prior Beliefs x Likelihood of Observed Event
and remember to keep it simple!
What is Occam’s Razor?
hint: william of occam
“the simplest explanation(that fits the observed data) is probably the correct explanation.”
People prefer explanations that explain more data with a minimal number of assumptions
What are explanations?
They are hypotheses for what we expect to occur!
What does Baye’s rule suggest?
Explanations should be combined with the data to update our belief in the hypothesis
Are all explanations created equal?
What might influence them?
No! our prior beliefs can influence our explanations
Backwards blocking characteristic.
Both on, One on, Other?
Off
Indirect Screening Off characteristic
Both on, one off, both
Both ON
Backwards blocking with prior examples.
Previously, backwards blocking was both on, one on, other OFF
1. BUT what happens if we see a lot of marbles already get through as blickets? does this influence the study
2. What about cubes, if we see all cubes as previously not blickets, what happens than? remember, normally backwards blocking is both on, one on, other off. what about blocks
- yes, we see both as blickets
2. we see both as not blickets