Slides - Andrew.cmu.edu

advertisement
Unit 5:
Activation and Latency
February 11, 2003
Subsymbolic Quantities
 Symbolic processing: chunks and productions
 Subsymbolic processing:
– chunk activation: controls retrieval
– production utility: controls conflict resolution
Unit 5
2
Activation
The activation of a chunk depends on:

History of usage of the chunks (how
frequently and recently the chunk has been
used)  base-level activation

Context in which the chunk is used 
spreading activation
Unit 5
3
Activation
Activation of chunk i
Base level activation of chunk i
Spreading activation
Activation from the source of activation j
(typically 1/n )
Strength of association between j and i
Unit 5
4
Chunk Retrieval
(p request-retrieval
=goal>
ISA geography-test
country =x
capital =c
city nil
state “get-retrieval”

=goal>
state “retrieving”
+retrieval>
ISA city-fact
country =x)
(p get-retrieval
=goal>
ISA geography-test
state “retrieving”
country =x
city nil
=retrieval>
ISA city-fact
city =y

=goal>
state “done”
city =y)
Unit 5
5
Situation at Time of Retrieval
Goal
ISA city-fact
country France
city Paris
Chunks
ISA city-fact
country Spain
city Barcelona
ISA geography-test
country France
city nil
capital Paris
Goals bias certain chunks
ISA city-fact
country France
city Marseille
- Which chunk should be
retrieved?
- How long does it take?
ISA city-fact
country France
city Orleans
Both answers depend on the
concept of activation.
Unit 5
6
Spreading Activation
ISA city-fact
country France
city Paris
Goal
ISA city-fact
Country Spain
city Barcelona
ISA geography-test
country France
city nil
state “retrieving”
capital Paris
S11
ISA city-fact
country France
city Marseille
France
S12
ISA city-fact
country France
city Orleans
S13
Unit 5
S21
Paris
7
Spreading Activation
ISA city-fact
country France
city Paris
ISA city-fact
Country Spain
city Barcelona
W2S21
ISA city-fact
country France
city Marseille
ISA city-fact
country France
city Orleans
Goal
W1S11
ISA geography-test
Source 1
country France
city nil
state “retrieving”
capital Paris
Source 2
WjSji
W1S12
France
j sources
i chunks
W1S13
Unit 5
Paris
W1…n = 1/n
Where n is the spread from goal
8
Spread of Activation From Source
(AKA Fan)
ISA city-fact
country France
city Paris
S1
ISA city-fact
country France
city Marseille
ISA city-fact
country France
city Orleans
fan=n+2
France
S2
S3
S5
S4
ISA geography-test
country France
city nil
Unit 5
9
Why it Makes Sense
 The source of activation contributes less (W) as the
number of goal source slots increases (Wj=1/n)
– This can be though of as a limit on how much can be profitably
focused on at once
– As more chunks are in goals slots, so attention (working memory)
gets spread thin
 Spreading activation also is less if it has to spread to
more things… S-ln(n)
– As n increases the strength of the spread decreases
*note the small n’s refer to different quantities
n in the first case is number of goal sources
n in the second case is the number of chunks matching the source
Unit 5
10
Two limits
 Limitation on spread as a function of how many
items W spreads to
– (our class model will not utilize this)
 Limitation on spread as a function of the number
of chunks that the Wj source spreads to
– S-ln(n), where n is the number of other chunks
Unit 5
11
A Retrieval Situation
two slots in goal,
so w=1/2
say bike is in 2 chunks
so bike WS=.5*(S-ln(4))
say umbrella is in 3 chunks
so umbrella WS=.5*(S-ln(5))
+Retrieval>
isa x
slot1 bike
slot2 helmet
slot3 =item
=Goal>
isa q
slot5 bike
slot6 umbrella
only bike is relevant
source currently
spreading so only bike
applies to retrieval
Chunk1
isa x
slot1 bike*
slot2 helmet
slot3 coat
Chunk4
isa q
slot5 umbrella*
slot6 disco
Chunk7
isa n
slota umbrella*
slotb classic
slotc classical
so, for our retrieval
A=B+WS=.5*(S-ln(4))
+retrievals get
the highest
activation chunk
Even non-matching chunks can be
retrieved if they are highest
after penalty (in this lesson nonmatching chunks will only be
retrieved if the correct chunk is
not present because there is no
noise in the equation)
Unit 5
Chunk2
isa x
slota bike*
slotb rock
slotc punk
Chunk3
isa z
slota badmusic
slotb umbrella*
slotc rap
Chunk5
isa q
slot5 7
slot6 banana
Chunk6
is m
slota bear
slotb lion
slotc tiger
* places receiving
activation boost
Can you ever retrieve
a chunk of the wrong
type?
12
Retrieval Latency
Time to retrieve chunk i
Latency factor
Activation of chunk i
Unit 5
13
The Sternberg Task
Study
Test
3
9
6
6
Unit 5
14
The Sternberg Task
Unit 5
15
Sternberg-Task Model
ISA do-Sternberg
list list
item “3”
State study
3 item list
ISA do-Sternberg
list list
item “9”
State study
Sji
Sji
ISA do-Sternberg
list list
item “6”
State study
Sji
List
j =1 (source spread from the list context)
Sji = S – ln (n+2)
n increases  Sji decreases 
Ai decreases Latency increases
Unit 5
ISA do-Sternberg
list list
State “attend-test”
item “3”
Goal
16
Sternberg Task: Targets
(sgp :lf 0.64) ; sets F
(setq *s* 3) ; sets S
Unit 5
17
The Fan Effect (Anderson, 1974)
Study
A <person> is in the <location>
<person> could have occurred in 1, 2, or 3 facts.
<location> could have occurred in 1, 2, or 3 facts.
A hippie is in the park.
A hippie is in the church.
A hippie is in the bank.
A captain is in the park.
A captain is in the church.
A debutante is in the bank.
A fireman is in the park.
Test: studied?
A hippie is in the park. (target)
A debutante is in the church. (foil)
Unit 5
18
Fan-Experiment Results
Targets
1.4
person
fan 1
person
fan 2
person
fan 3
1.35
1.3
RT
1.25
1.2
1.15
1.1
1.05
1
loc fan 1
loc fan 2
RT
Foils
1.5
1.45
1.4
1.35
1.3
1.25
1.2
1.15
1.1
1.05
1
loc fan 1
loc fan 3
Unit 5
loc fan 2
loc fan 3
19
Tips
 Read only the person and the location
The lawyer is in the church
Use =visual-location>
… screen-x (within 10 20)
 Meanings
(lawyer* isa meaning word “lawyer”)
 Two productions that retrieve a fact
 Will need to find S and F that best fit data
Unit 5
20
Download