2. How does working memory work ?
• How is the relevant information maintained active
during processing ?
• What is the nature of the resource to be shared ?
• How is this sharing achieved ?
• What are the limiting factors of WM functioning ?
3. Time-Based Resource-Sharing Model
The main proposals
Barrouillet, Bernardin, & Camos, JEP:G, 2004
1. Processing and maintenance require attention which is a limited
resource (some sharing is needed)
2. As soon as attention is switched away, activation suffers from a
time-related decay
3. Among attention demanding activities, retrievals from memory
should have the most detrimental effect on concurrent maintenance
4. When processing involves retrievals, sharing attention is time
based because a central bottleneck allows only one retrieval at a
time
4. 1. Processing and storage require attention
Anderson’s ACT-R Framework
Processing Maintenance
Productions rules read and
update the content of WM STM = Activated part of LTM
Retrievals of declarative Frequent refreshment of
knowledge stored in long-term decaying traces of the to-be-
memory maintained items
Activation is of memory items
Activation attention demanding
5. 2. When attention is switched away, activation
suffers from a time-related decay
• Activation is produced by attentional focusing (Cowan, 1995).
• Activation declines as soon as the focus of attention is
switched away.
• While processing captures attention, relevant information
declines in STM
• When attention is used to refresh decaying memory traces,
processing is temporarily suspended.
6. 3. Concurrent memory retrievals have the
most detrimental effect on maintenance
• The refreshment of the decaying memory traces in STM
necessitates their memory retrieval, but
• Two memory retrievals can not be performed simultaneously
(Pashler, 1998; Rohrer, Pashler, Etchegarray, 1998).
Any processing component that requires retrievals from
memory should have a highly detrimental effect on
concurrent maintenance of information.
7. 4. Sharing attention is time-based
• There is a bottleneck for retrievals: only one retrieval at a
time.
• Maintenance necessitates frequent retrievals
• When processing occupies the bottleneck …
Big problem
9. Switching mechanism and decay
Possible reactivation
of memory traces CL
Rabbit R R R R Diner
CL
Rabbit R R R R Diner
CL
Rabbit R R R R Diner
10. Cognitive Load is
The proportion of time during which a given activity captures attention
in such a way that the refreshment of memory traces is impeded.
Duration of attentional capture
CL =
Total time allowed
The higher the cognitive load, the more difficult the switching.
11. A metric for Cognitive Load
In tasks involving retrievals from LTM
The number of retrievals n
Their difficulty a
(the time they occupy central processes)
The total time allowed to perform them T
When all the retrievals are identical in
nature:
Σ ai ni aN
CL = CL =
T T
12. Exploring cognitive load as the
Number of Retrievals / Time ratio
The Reading Digit Span Task
R8
31 Read aloud the successive screens
64 and recall the letters
K7
25
49
L3
68
24
13. Manipulating the
Number of Retrievals / Time ratio
Either 6 or 10 digits to be read
Constant duration of the interletter intervals (6 s)
4,5
4
3,5
3
2,5
2
1,5
1
0,5
0
6 Digits 10 Digits
14. Manipulating the
Number of Retrievals / Time ratio
Fixed number of digits to be read
Either 600 or 1000 ms per digit
5
4,5
4
3,5
3
2,5
2
1,5
1
0,5
0
Slow Fast
1000 ms 600 ms
15. Manipulating the
Number of Retrievals / Time ratio
Varying
the number of digits to be read
and the time allowed to read them
•Either 4, 8, or 12 digits during 6, 8, or 10 seconds
• 9 different values of the critical ratio (from 0.4 to 2)
16. Manipulating the
Number of Retrievals / Time ratio
6
5,5
5
4,5
4
3,5
3
2,5
R2 = .932
2
1,5
1
0 0,5 1 1,5 2 2,5
Number of retrievals / Time ratio
Barrouillet, Bernardin, & Camos, JEP:G, 2004
17. A physical law for mental effort
Cognitive Load is
Work
CL =
Time
The physical law of power !
Cognitive Load of a given activity = Mental Power needed to perform it
18. Cognitive Load
as defined by the Time-Based Resource-Sharing model
depends on
rate of processing rather than complexity
nature of the processes involved
attentional demand of the processes
duration of the atomic steps of processing
19. Rate of processing rather than complexity
Lépine, Bernardin, & Barrouillet, EJCP, in press
In undergraduate students who remembered series of
letters:
• Traditional Reading Span (self paced)
• Reading Letter Span (slow: 1200 ms per letter)
• Reading Letter Span (fast: 600 ms per letter)
20. Rate of processing rather than complexity
Lépine, Bernardin, & Barrouillet, EJCP, in press
4,5
4
3,5
3
2,5
2
1,5
1
RS self-paced RLS slow RLS fast
Reading letters can have the same detrimental effect on spans as reading complex sentences !
21. Nature of the processes involved
Bernardin, Portrat, & Barrouillet, submitted
Two different groups are presented with the same display
G 8 but perform different activities:
5
6
location
Parity 1 2 “ Up, up, down, down”
“ Even, odd, even, odd …”
Retrievals from LTM required
3
P
22. Nature of the processes involved
Bernardin, Portrat, & Barrouillet, submitted
Involvement of central processes or tracking the target ?
G8 G8
Chao cti ai
prese n t t on
5
5
6 6
1 2 1
Regular presentation
2
3
P 3
P
23. Nature of the processes involved
Bernardin, Portrat, & Barrouillet, submitted
4,5
ns
4
3,5
*
3
2,5
Regular
2 Chaotic
1,5
1
0,5
0
Position Parity
Tracking external events is far less demanding than occupation of central processes
Cognitive load mainly results from central processes occupation
24. Attentional demand of the processes
Gavens & Barrouillet, JML, in press
R5 R1 Ordered
31 23
24 45
K3 K1
25 23
Random 4L 4L
43 12
65 34
21 56
25. Attentional demand of the processes
Gavens & Barrouillet, JML, in press
3
2
Ordered
Random
1
0
8-year olds 10-year olds
26. Duration of the processes
aN Slower retrievals
CL =
T
Central processes occupied for a longer period
Higher CL
LOWER SPANS
27. Duration of the processes
A reading digit span with digits presented …
4 Four IV
442 ms 446 ms 625 ms
Reading digit spans should be lower when digits are presented in roman
Reading numbers (1 to 9) while maintaining letters
1 digit per second
28. Duration of the processes
4,5
4
3,5
*
3
2,5
2
4 Four IV
Slower retrievals occupy central processes for longer periods
and involve higher cognitive load.
29. Are time-constrained tasks as predictive as
classical WM span tasks?
Lépine, Barrouillet, & Camos, Psych.B&R, in press
Predicting academic achievement in 93 sixth graders
Working memory Academic achievement
Classical tasks
Operation span .36
Reading span .34
Compound score .39 Standard National
evaluation in litteracy
New tasks and mathematics
Continuous Operation span .42
Reading Letter span .50
Compound score .54
30. The Time-Based Resource-Sharing model
and the new WM span tasks
More predictive
Controling time Simpler
parameters provides us
with better WM span Easier to manipulate
tasks Easier to control
More knowledge free
31. Conclusions
The main function of WM is to share cognitive resources between maintenance and
treatment: Perfect trade-off.
The more constrained this sharing, the higher the cognitive load: What matters is PACE !
Is there an intrinsic cognitive load for a given task ?: No, just an amount of work to be
done
Any task that involves central processes can become very demanding when performed
under sufficient time pressure.
Demanding activities in traditional cognitive psychology ?: Activities for which time
pressure is inherent to their structure.
Cognitive Load is not a myth or even a metaphor.
Cognitive Load is:
Proportion of time attention is totally captured
Work to be done / Time to do it
Mental power needed
32. Thanks to
Sophie Bernardin
Raphaëlle Lépine
Nathalie Gavens
Sophie Portrat
LEAD - CNRS Université de Bourgogne
Notes de l'éditeur
However, little is known about the precise mechanisms that underlie WM functioning, and many questions remain unanswered.
Our model is aimed to answer these questions. It is based on four main proposals
As far as the first proposal is concerned, we assume that processing and storage both require attention. Within the ACT-R framework, processing most often require to retrieve knowledge from long-term memory, e.g., for reading, or solving operations. The storage component requires to regularly retrieve the items of relevant information because their memory traces suffer from a decay and tend to fade away. Thus, both activities require retrievals from memory that require attentional focusing and thus attention
Second, we assume that the activation of memory items is produced by attentional focusing, as suggested by many models of WM such as Anderson ’ s ACT-R or Cowan ’ s model. Thus, when the focus of attention is switched away, activation declines with time.
Third, According to Pashler, there is a central bottleneck that constrains retrievals. Thus, it would be impossible to perform simultaneously two memory retrievals. The retrievals required by processing impede the refreshment of ST memory traces. As a consequence …
Forth, sharing attention is time-based. because there is a central bottleneck, the memory retrievals needed by the processing component captures attention and, as a consequence, the memory traces decay and vanish. We have a problem
Thus, The solution is to rapidly switch between processing and storage.
We assume that the constraints on the switching process determines cognitive load. Suppose that your are performing a WM span task, and that you are successively presented with these two words. Unfortunately, you have to perform some intervening task during the interval. This task involves memory retrievals, during which the memory traces of the words decay, but you can keep free short slots to retrieve and refresh these traces. Let us suppose that this activity involves a moderate cognitive load. Now, if you are given more time to perform the same task, you have longer periods of time to reactivate memory traces. The WM task becomes easier/ because cognitive load decreases On the other hand, if the time is reduced, it becomes difficult to concurrently refresh memory traces. Cognitive load increases /and WM span should decrease.
CL depends on the number and difficulty of retrievals because retrievals, among other processes, block attention for a portion of time
According to this analysis, the CL of a task that mainly involves memory retrievals depends on three parameters: the number of retrievals it requires, their difficulty, (some retrievals need more time than others), and the time allowed to perform them CL is given by the following equation, Which can be simplified when all the retrievals are identical in nature and difficulty. CL depends on the Number of retrievals / Time ratio .
Our second hypothesis concerned the effect of time on cognitive load. Thus we tested our hypothesis that CL is a function of the number of retrievals / Time ratio We used a very simple task, the reading digit span task letters to be remembered and digits to be read are successively presented on screen And subjects are asked to read them aloud and to recall the letters.
As we predicted, increasing the number of digits to be read had a detrimental effect on span
Once more, there was a clear effect of pace. The faster the pace, the lower the span. Note that in this case, the task remained unchanged and its duration was reduced. Thus, the delays of retention were reduced. Towse and Hitch would predict higher spans for shorter delays. But shorter delays resulted in lower spans, that is absolutely at odds with their hypothesis.
We verified the predictive value of the equation in a study in which we varied both the number of digits to be read and the time allowed to read them.
As we predicted, WM spans decreased as CL of the reading digits tasks increased. As expected, the same phenomenon was observed in the baba span, but of course less pronounced
What is CL? CL is in fact a function of the cognitive work to be done, divided by the time to do it. This function corresponds to the physical law of power. Cognitive load corresponds to the mental power needed to perform an activity. This means that simple tasks can result in very high cognitive load.
Recall that our theory predicts that very simple tasks can have a highly detrimental effect on spans because complexity does not matter. What is important is pace. Actually, our experiments support this prediction. We can compare three tasks in which undergraduate students had to recall letters: The operation span that involves a quite complex activity of problem solving The continuous operation span in which adults have just to solve very elementary operations like adding or subtracting 1 or 2 And the reading digit span in which adults have just to read digits, a very simple activity.
The mean reading digit span was about 3 Not surprisingly, continuous operation span is slightly lower than the reading digit span, But operation span is better! Lire l ’ écran
We predicted also that among attention-demanding activities, the retrievals should have the most detrimental effect because retrievals are also needed to refresh decaying memory traces in STM. We designed a task … Our prediction is that, all other thinks being equal, the parity task would result in lower spans than the decision about location.
In the same experiment, we compared spans from parity and location tasks in which the targets were not presented at a regular and smooth rhythm but at a Chaotic rhythm
As we expected, an activity that requires retrievals from LTM is more disruptive than an activity that just requires response selection. By contrast, the way the targets are presented had only a small and non significant effect.
If our analysis is correct, the time a given activity blocks the central processes has a direct impact on the CL. This is the parameter a in the CL equation. For example, slower retrievals occupy the central processes for longer periods of time, Thus involve a higher CL and then lower spans.
We compared three conditions of a reading digit span in which numbers were presented either in their Arabic Verbal or Roman form. The corresponding Reading times are … Thus we predicted lower spans when numbers are presented in their Roman form.
As predicted by the reading times, there was no big difference between spans when digits were presented in Arabic or Verbal form. However, the roman presentation, which takes longer to read resulted in lower spans. As preicted by the Time-based resource sharing model …
The main interest of WM span tasks is their predictive value. Are the new tasks as predictive as the classical operation or reading span tasks? We investigated this question by evaluating the correlation between WM spans and academic achievement in 93 sixth graders. We used classical … as well as new WM span tasks. The reading letter span task is a task ni which children had to read letters on successive screens while remembering digits. Academic achievement was evaluated using the score on the national tests. The venerable operation span made a good job with a correlation of .36 But continuous operation span was slightly better. Daneman and Carpenter ’ s reading span was not so bad, but the reading letter span was far more predictive. In fact, when we compared the compound scores, the classical tasks correlated with academic achievement, but the new tasks were more predictive.
In fact, time-constrained tasks have many advantadges. They are more predictive, but also they are simpler They are easier to manipulate. For example, we can easily manipulate the duration of the task, its difficulty They are easier to control, because they are computer-paced and you can control subject ’ s activity And finally, they are more knowledge free. They involve simple activities that everybody is able to complete. Some sixth graders are probably poor readers, or mathematically disabled, but they are all able to read letters or digits. Thus low spans can not be due to the fact that the subject is unable to correctly perform the secondary task.
Some provisional conclusions from a work in progess.