Working Memory Updating Decomposed

1
Working Memory Updating Decomposed

• Ullrich Ecker1, Stephan Lewandowsky1, Klaus
  Oberauer2, Abby Chee1
• 1 University of Western Australia, 2 University
  of Bristol

2
Working memory updating (WMU)

• Working memory (WM) holds selected
  representations available for ongoing processing
• To maintain accurate representations of
  information that changes over time, WM content
  needs to be updated
• Example of WMU Keeping score in a tabletennis
  match
• Important for
• Mental arithmetic
• Language comprehension
• Navigating through traffic
• etc.

3
Aims of study (1)

• WMU not a unitary process
• WMU needs to allow for stability and flexibility
  at the same time (Kessler Meiran, in press)
• If X is replaced with Y
• and then Z
• Unclear
• What are the component processes of WMU?

4
Aims of study (2)

• Individual differences research
• WMU specifically predicts higher cognitive
  abilities
• However, WMU often measured with tasks that
  conflate WMU WM capacity (WMC)
• Running memory task U O N S A C
• Unclear
• What is the relationship between WMU WMC?
• WMU WMC may rely on common WM abilities
  (Schmiedek et al., in press)
• WMU WMC may be distinct and dissociable (van
  Raalten et al., 2008 Radvansky Dijkstra, 2007)
  
• Predictive power of WMU over and above WMC?

5
Aims of study (2)

• Individual differences research
• WMU specifically predicts higher cognitive
  abilities
• However, WMU often measured with tasks that
  conflate WMU WM capacity (WMC)
• Running memory task U O N S A C E
• Unclear
• What is the relationship between WMU WMC?
• WMU WMC may rely on common WM abilities
  (Schmiedek et al., in press)
• WMU WMC may be distinct and dissociable (van
  Raalten et al., 2008 Radvansky Dijkstra, 2007)
  
• Predictive power of WMU over and above WMC?

6
The present study

• Decomposition of WMU into three distinct
  components
• Retrieval
• Transformation
• Substitution
• Implemented into a standard WMU paradigm
• Plus WMC battery
• 97 subjects

7
The present study

• Decomposition of WMU into three distinct
  components
• Retrieval
• Transformation
• Substitution
• Implemented into a standard WMU paradigm
• Plus WMC battery
• 97 subjects

actually, 5 more than the 20 initially expected
8
The present study

• Decomposition of WMU into three distinct
  components
• Retrieval
• Transformation
• Substitution
• Implemented into a standard WMU paradigm
• Plus WMC battery
• 97 subjects

actually, 5 more than initially expected
9
The present study

• Decomposition of WMU into three distinct
  components
• Retrieval
• Transformation
• Substitution
• Implemented into a standard WMU paradigm
• Plus WMC battery
• 97 subjects

actually, were expecting 25
10
The present study

• Decomposition of WMU into three distinct
  components
• Retrieval
• Transformation
• Substitution
• Implemented into a standard WMU paradigm
• Plus WMC battery
• 97 subjects

actually, 4 people cancelled, but 2 new couples
11
The present study

• Decomposition of WMU into three distinct
  components
• Retrieval
• Transformation
• Substitution
• Implemented into a standard WMU paradigm
• Plus WMC battery
• 97 subjects

12
Updating task Trial structure

• Encoding
• Remember 3 letters-in-frames
• Updating
• 6 steps
• Update individual frames (alphabet arithmetic)
• Remember and type result (? RT, accuracy)
• Finall recall (not analysed)
• To make sure subjects updated until the end,
  given predictable sequence
• All frames in random order

13
Design

• Updating
• 3 factors Retrieval, Transformation, Substitution
• fully crossed within-subjects

14
Updating conditions

• Pure Substitution
• type N, remember N F A

15
Updating conditions

• Pure Transformation
• type F, remember V F A

16
Updating conditions

• Pure Retrieval
• type A, remember V F A

17
Updating conditions

• All 3 combined
• type H, remember V H A

?2
18
Updating conditions

• Baseline condition
• type V, remember V F A

V
19
Updating conditions

• Retrieval plus Substitution
• type F, remember V F F

20
Updating conditions

• Transformation plus Retrieval
• type A, remember V F A

?0
21
Condition prompts at a glance
Assuming C is currently remembered
22
Sample trial

Im ready
23
Sample trial Initial encoding
B-Q-J
Encoding time 2 s
24
Sample trial Updating step 1
?1
C-Q-J
25
Sample trial Updating step 2
I2
C-Q-K
Note frame switch on every step
26
Sample trial Updating step 3
?
C-Q-K
Note frame switch on every step
27
Sample trial Updating step 4
S
S-Q-K
Note frame switch on every step
28
Sample trial Updating step 5
S-K-K
Note frame switch on every step
29
Sample trial Updating step 6
J1
S-K-K
Note frame switch on every step
30
Sample trial Final recall 1
?
K
Random order
31
Sample trial Final recall 2
?
S
Random order
32
Sample trial Final recall 3
?
K
Random order
33
Results
2 MODELS
34
Modelling

• Multilevel regression SEM
• Starting point
• Maximal parsimony Could there be additivity?

35
Multilevel regression (RT)

• Additional frame-switch in ? condition
• frame-switch cost (Garavan, 1998 483 ms) was
  subtracted from all ? RTs
• Transformations were represented by 3 dummy
  variables
• T1 1 (baseline)
• T2 2 (coded as RT increase relative to
  baseline T1)
• T0 0 (coded as RT decrease relative to
  baseline T1)
• UpdRT 1.30 .04R 1.20T1 .94T0 .58T2
  .30S
• Coefficient of discrimination r2(obs, fitted)
  across all 1197 data points .91
• Likelihood ratio tests Fit not improved by
  adding interactions ? additivity

36
SEM
e12
OS
e11
SS
WMC
e10
SSTM
e9
MU
37
SEMRT model
Strong constraints that impose additive
structure All loadings fixed to 1 all manifest
intercepts and error means fixed to 0
RT.1
e1
S
RT.2
e2
RT.3
e3
T
e4
RT.4
RT.5
e5
R
RT.6
e6
e7
RT.7
GenRT
e8
RT.8
38
SEMRT model
Relaxed additivity constraints Fixed e3 to
frame-switch estimate (Garavan, 1998) freely
estimated TRT.5 weight allowed error covariation
RT.1
e1
S
RT.2
e2
e3
RT.3
T
.48
e4
RT.4
RT.5
e5
R
Condition 3 ? (2nd frame switch) Condition 5 ?0
RT.6
e6
e7
RT.7
GenRT
e8
RT.8
39
SEMRT model
?2(55) 94.02 CFI .94 RMSEA .09 SRMR .07
RT.1
e1
S
.34
RT.2
e2
e12
OS
e3
RT.3
T
-.37
.41
.48
.52
1.47
e4
e11
SS
RT.4
WMC
RT.5
e5
R
e10
SSTM
.50
.03
RT.6
e6
.36
e7
e9
MU
RT.7
GenRT
-.50
1.28
e8
RT.8
40
SEMRT model
Methodological advancement Simultaneous
estimation of latent weights and means
RT.1
e1
S
.34
RT.2
e2
e12
OS
e3
RT.3
T
-.37
.41
.48
.52
1.47
e4
e11
SS
RT.4
WMC
RT.5
e5
R
e10
SSTM
.50
.03
RT.6
e6
.36
e7
e9
MU
RT.7
GenRT
-.50
1.28
e8
RT.8
41
Estimated ? observed means

• Estimated means can be used to accurately
  re-calculate observed experimental data
• For example, in case of T S
• 1.28 1.47 .34 3.08 3.01 s
• Median deviation 35 ms

42
SEMAcc model
Strong constraints that impose additive
structure All loadings fixed to -1 all
manifest intercepts and error means fixed to 0
43
SEMAcc model
Relax additivity constraints Freely estimated
TAccpb.5 weight and e4e8
Accpb.1
e1
S
Accpb.2
e2
e3
Accpb.3
T
Accpb.4
e4
Accpb.5
e5
R
Accpb.6
e6
Condition 5 0 Assuming C is
remembered Condition 4 X Condition 8 C
e7
Accpb.7
GenAcc
Accpb.8
e8
44
SEMAcc model
?2(56) 75.77 CFI .95 RMSEA .06 SRMR .08
Accpb.1
S
.17
Accpb.2
e12
OS
Accpb.3
T
-.38
.36
.24
e11
SS
Accpb.4
WMC
-.72
Accpb.5
R
e10
SSTM
.41
Accpb.6
-.44
e9
MU
Accpb.7
GenAcc
.38
1.62
Accpb.8
45
Estimated ? observed means

• Estimated means can be used to accurately
  re-calculate observed experimental data
• For example, in case of T S
• 1. 62 - .24 - .17 1.21 (probit) ? .89 .87
• Mean deviation lt .03 accuracy units

46
Summary

• Data suggest that R, T, and S
• make orthogonal and additive contributions to WMU
• probably run serially R?T?S
• Transformation
• strong effect on both WMU accuracy and RT
• covaries with WMC
• No actual retrieval process operating (no effect
  on RT)
• subjects likely keep all 3 letters in a region
  of direct access (Oberauer, 2002)
• Accuracy effect of retrieval
• direct access vs. integrity / accuracy of
  representation
• covaries with WMC
• Substitution
• relatively small but significant effect on both
  WMU accuracy and RT
• the only factor with own variance separate from
  WMC

47
Conclusion

• Only substitution is unique to WMU.
• Previous reports that WMU specifically predicts
  higher cognitive abilities are likely due to a
  conflation of WMU and WMC factors.
• Studies that reported dissociation of WMU and WMC
  (e.g., in schizophrenia, van Raalten et al.,
  2008) used WMU tasks that relied heavily on
  substitution.
• To-be-tested Are substitution skills in
  themselves useful in predicting higher cognitive
  abilities?

48
Thank you!

• References
• Garavan, H. (1998). Serial attention within
  working memory. Memory Cognition, 26, 263-276.
• Oberauer, K. (2002). Access to information in
  working memory Exploring the focus of attention.
  Journal of Experimental Psychology Learning,
  Memory, and Cognition, 28, 411-421.
• Radvansky, G. A., Dijkstra, K. (2007). Aging
  and situation model processing. Psychonomic
  Bulletin Review, 14, 1027-1042.
• Schmiedek, F., Hildebrandt, A., Lovden, M.,
  Wilhelm, O., Lindenberger, U. (in press).
  Complex span versus updating tasks of working
  memory The gap is not that deep. Journal of
  Experimental Psychology Learning, Memory, and
  Cognition.
• Van Raalten, T. R., Ramsey, N. F., Jansma, J. M.,
  Jager, G., Kahn, R. S. (2008). Automatization
  and working memory capacity in schizophrenia.
  Schizophrenia Research, 100, 161-171.
• Thanks to

Steve Lewandowski Klaus Oberauer
Abby Chee