T h e S u ff ic ie n t F ea tu re M o d el A p p li ed t o t h e A u to m at iz at io n o f a V is u al a n d M em o ry S ea rc h T as k
By Denis cousineau Cousined@PSY.UMontreal.CAO v er v ie w o f th e P re se n ta ti o n
1- Basic findings in the memory and visual search task 2- Overview of the Sufficient Feature Model (SFM) 3- Results of a Computer Simulation 4- Further Results 5- Future Direction of Research1 - B as ic F in d in g s in t h e M em o ry an d V is u al S ea rc h T as k
•Two loads are manipulated: Memory load and Display load •We measure the reaction time; subjects must keep their error rates lower than 5%LHBZ H ·*The answer is: “Present”
•In consistent mapping conditions (CM), the load effect vanishes with practice: •In varied mapping conditions (VM), the learning is small, the load effect remains large: 124 Display size Rea
cti on ti mes
M=1M=2M=4
124 Display size
Rea cti on ti mes
M=1M=2M=4
2 - T h e S u ff ic ie n t F ea tu re M o d el (S F M ): B as ic A ss u m p ti o n s
The SFM has three basic assumptions: 1) comparison mechanism: based on non-atomic representation of stimuli (feature representation) 2) learning mechanism: based on the reduction of information using: •Ordering of the feature tests based on diagnosticity •Reducing the number of feature tests performed 3) limited-capacity handling of features (an attentional window)D ia g ra m o f th e D ec is io n M ec h an is m
HLRLB ZGMemory LoadDisplay Load (11, 6, 2, 7), (11, 2), (5, 2, 1, 7) (11, 2), (11, 1, 9), (2, 1), (2, 1, 7) Attentional Window
11, 5, 6, 2, 1, 9, 7, 3, 8, 4, 10, 12
Feature Test list
+ -
diagnosticity Useless features
T h e S u ff ic ie n t F ea tu re M o d el (S F M ): P ar am et er s
The SFM has three parameters: –%Error accepted: with 5%, it accepts 1 error out of 20 trials –Size of the attentional window: it must be large enough to contain the features of the most complex character –Character composition: we used simplified characters, composed of lines onlyC h ar ac te rs U se d i n t h e S im u la ti o n
Inspired by McClelland and Rumelhart (1981) The character “2” is described by the list (1, 4, 6, 8, 12)T h e S u ff ic ie n t F ea tu re M o d el (S F M ): C o m p u te r S im u la ti o n
•The SFM has no free numerical parameters •It measures the number of features that went under the attentional windows •The simulation is trained by practicing on the same trials as human did •There is no forgetting in the simulation3 - B eh av io r o f th e S F M
Results of the simulation are in this order: •Learning curves data •Mean number of features considered •Standard deviation of the number of features considered They are compared to two groups of 4 human subjects, one trained for 12 blocks in CM condition, the other trained in VM conditionL ea rn in g C u rv e
The ordering of the feature tests results in a faster rejection of distractors Therefore, there is a visible learning curve both in CM and in VM conditionsCM 0
100
200
300
400
500
600
700
800
900
1000 123456789101112 Block
Mill ise co nd s
VM 0
100
200
300
400
500
600
700
800
900
1000 123456789101112 Block
Mill ise co nd s
CM 02468
10
12
14
16
18
20
22
24
26
28
30 123456789101112 Block
Nu mb er o f c om pa ris on s
RT pos RT neg SD pos SD neg VM 024681012141618202224262830 123456789101112 Block
Nu mb er of co mp ari so ns
Learning rates are equalsLearning rates are also equals
M ea n R ea ct io n T im e
•In CM condition, a limited number of features discriminates all targets from distractors. There is no overload of the attentional windows •In VM condition, the number of features needed remains large. The attentional windows must swap from character to character200300400500600700800900100011001200 1234 D size
mea n R T
0%10%20%30%40%50%60%70%80%90%100%
Percent Error
200
300400
500
600
700
800900
1000
1100
1200 1234 D size
mea n R T
0%
10%20%
30%
40%
50%
60%70%
80%
90%
100%
Percent Error
010
20
30
40
50
60
70 1234 D size
N o f f ea tu re s
010
20
30
4050
6070 1234 D size
N o f f ea tu re s
200300400500600
700
8009001000
11001200 1234 D size
mea n R T
0%10%20%30%40%
50%
60%70%80%
90%100%
Percent Error
200300
400500
600700
800900
100011001200 1234 D size
mea n R T
0%10%
20%30%
40%50%
60%70%
80%90%100%
Percent Error
010
20
3040
50
60
70 1234 D size
N o f f ea tu re
0 s
10
20
30
40
50
60
70 1234 D size
N o f f ea tu re s
S ta n d ar d D ev ia ti o n
In the SFM: –Negative trials: To be exhaustive on character locations means to be self-terminating on features –Positive trials: To be self-terminating on character locations means to be exhaustive on features located above diagnostic threshold Therefore, variance on positive and negative trials should behave in a similar fashion. It is opposed to a strict serial model, where negative trials should have a smaller variance010
20
30
40
50
60
70 1234 D size
Sta nd ard D evi atio n
010
20
30
40
50
60
70 1234 D size
Sta nd ard D evi atio n
050100150200250300350
400450500 1234 D size
Sta nd ard D evi ati on
050100150200250300350400450500 1234 D size
Sta nd ard D ev iat io n
050100150200250300350400450500 1234 D size
Sta nd ard D ev iat io n
050100150200250300350400450500 1234 D size
Sta nd ard D evi ati on
010
20
30
40
50
60
70 1234 D size
Sta nd ard D ev iat io n
010
20
30
40
50
60
70 1234 D size
Sta nd ard D ev iat io n
S u m m ar y o f th e R es u lt s
A decision process based on reduction of information assumptions reproduces very well the means and standard deviation for both CM and VM conditions. This is done by: –evaluating diagnosticity on a trial by trial basis and by –testing for useless features, based on the % of error allowed.A) Categorical Manipulation In a CM conditions, targets are always the same. They can be seen as a category. Do pre-established categories alter performance?
4 - F u rt h er R es u lt s
We tested CM conditions with targets and distractors being of the same category of characters (HOMO condition) or of distinct categories (HETERO condition). Example of setstargets / distractors HOMOL, H, R, S / Z, B, G, F vs HETERO2, 3, 6, 7 / Z, B, G, F
CM HETERO group achieve asymptote on the second session of learning. They show no load effect on the first block CM HOMO never equals CM HETERO with more than 5000 trials of practice. CM HETERO 0
100
200
300
400
500
600
700
800
900
1000 123456789101112 Block
CM HOMO 0
100
200
300
400
500
600
700
800
900
1000 123456789101112 Block
Mil lis ec on ds
With CM learning, a generic representation (GR) composed of diagnostic features of a category might become easier to retrieve than representation of a single target (T). This raises the possibility that, after such an association is learned, it might be possible to performed well under Inconsistent Mapping (IM) conditions. One type of IM is seen next.
TGRTest ProbeDecision
B) Categorical Varied Mapping (CVM) conditions If GR is easy to retrieve, it might be possible for subjects to alternate between two distinct and mutually exclusive sets of targets and distractors Example of setstargets / distractors HOMOL, H, R, S / Z, B, G, F and Z, B, G, F / L, H, R, S HETERO2, 3, 6, 7 / Z, B, G, F and Z, B, G, F / 2, 3, 6, 7
•CVM HETERO results are comparable to CM results (learning curve and load effect) •CVM HOMO results show learning curve similar to VM groups and large load effect. CVM HOMO 0
100
200
300
400
500
600
700
800
900
1000 123456789101112 Block
Mil lis ec on ds
CVM HETERO 0
100
200
300
400
500
600
700
800
900
1000 123456789101112 Block
Mi lli se co nd s
010
20
30
40
50
60
70 1234 D size
N o f f ea tu re s
01020
30
40
5060
70 1234 D size
N o f f ea tu re
Model was given two feature s tests lists to learn from, one for letters, one for digits 1200 1100 1000 900 800 700 600 500 400 300 200 1234 D size
mea n R T
0%10%20%30%40%50%60%70%80%90%100%
Percent Error
200300400500600700800900100011001200 1234 D size
mea n R T
0%10%20%30%40%50%60%70%80%90%100%
Percent Error
Im p li ca ti o n o f th e S F M t o O th er T h eo ri es
A) Category knowledge: With a sensitivity to sets presented, the model can learn in the inconsistent CVM HETERO condition This is against any direct-access theory of performance (Logan’s Instance-based theory of automaticity for instance)B) Non-atomic representation: Physical similarity effects (such as Distractor- Distractor homogeneity effect and Target- Distractor heterogeneity effect) are easily explained by the SFM. Furthermore: With feature representation, standard deviation are well-described.
C) Reduction of information: It captures the essential aspects of CM and VM learning (load effect in both M and SD, some aspects of the learning curves)
The CM Character and VM Feature task (Work in progress with C. Lefebvre) Since characters have two levels of representation (holistic, featural), inconsistencies in mapping is possible:targets / distractors Will performance be more like VM or CM?
p d h u / b q n u v s p d b q / h u n u
5 - F u tu re D ir ec ti o n o f R es ea rc h : A n a p ri o ri t es t o f th e S F M
Thank you for your attention.
The GR can be seen as a decision tree where features are nodes, and tests are vertices. 11, 5, 2, 1, 6, 9, 7, 3, 8, 4, 10, 12
+ -
diagnosticity Useless features
11? 2?5? 1?
PresentAbsent Present « YES » (L, H, S) PresentAbsent « YES » (R)
« No »
Treisman reports linear variance; Ward and McClelland report linear STD; Cousineau and Larochelle find in- between results. SFM accepts a broad range of value
) ( ) ( ) ( ) ( ) ,1 (
2T m T m E Var Var E n Var + =
Number of feature mVa ria nce
m2 : Ward and McClelland SFM m: Triesman