Kognitive Architekturen

1
State of ACT-R Research Agenda

1. Review the ACT-R 5.0/6.0 architecture.
2. Illustrate its application to two experiments on
   learning to solve equations -- one with children
   and with adults.
3. Show how fMRI data provide converging data for
   architectural assumptions.
4. Show off nearly parameter-free predictions.
5. Discuss future goals.

2
ACT-R versus ACT

• Common Cognition has declarative procedural
  systems.
• Common Each system has subsymbolic symbolic
  aspects.
• ACT-R 2.0 Rational analysis guiding the
  subsymbolic level.
• ACT-R 4.0 Central cognition integrated with
  perceptual-motor.
• 5. ACT-R 4.0/5.0 Different types of learning
  that really work.

Knowledge Level Learning Module products
recorded as chunks -- focus on instruction and
examples
6. ACT-R 5.0 has biologically-inspired module
and buffer structure. 7. ACT-R 6.0 Use of this
structure to foster cumulative science.
3
ACT-R 5.0/6.0 Modules and Buffers
ACT-R
Parse 3x-57
Visual Perception
Manual Control
Type x4
Production System
Declarative Memory
Retrieve 7512
ProblemState
Hold 3x12
Control State
Unwinding Retrieving
4
The First Experiment Qin, Anderson, Silk,
Stenger, Carter (2004)

1. 11-14 year-olds just about to start Algebra 1
2. Day 0 Instruction, paper pencil practice,
   coaching
3. Days 1 - 5 Computer-based practice

4.Student types answer by pressing finger in data
glove. 5. Imaged in fMRI scanner on Day 1 and 5.
5
Unwind Instructions that ACT-R Parses into an
Internal Declarative Representation

• 1. To solve an equation, encode it and
• a. If the right side is a number then image that
  number as the result and then focus on the left
  side and unwind it.
• b. If the left side is a number then ..
• 2. To unwind
• a. If the expression is the variable then the
  result is the answer.
• b. If a number is on the right unwind-right
• c. If a number is on the left unwind-left
• 3. To unwind-right, encode the expression and
• a. If the expression is _ 0 then focus on the
  left part and unwind
• b. Otherwise invert the operator, image it as the
  operator in the result, image the right part of
  the expression as the second argument in the
  result, evaluate the result, and then focus on
  the left part and unwind
• 4. To unwind-left encode the expression and
• If the expression is 1 _ then focus on the
  right part and unwind
• Otherwise check that the operator is symmetric,
  invert the operator, image it as the operator in
  the result,

6
ACT-Rs General Procedures for Interpreting and
Following Declarative Representations of
Procedures applied to 7x 3 38

• Instruction 1a Create image 38
• Instruction 2b Unwind-right 7x3
• Instruction 3b Change image to 38-3, this to
  35, and focus on 7x
• Instruction 2c Unwind-left 7x
• Instruction 4b Change image to 35/7, this to
  5, and focus on x
• Instruction 2a The answer is 5, key it.
• Initially instructions are retrieved and
  interpreted.
• Eventually production compilation produces
  task-specific production rules.

7
Examples of Production Rules
General Interpretive If one has retrieved an
instruction for achieving a goal THEN retrieve
the first step of that instruction
Prior Knowledge IF one is evaluating the
expression a operator b THEN try to retrieve a
fact of the form a operator b ?
Acquired Task-Specific IF the goal is to unwind
an expression and the expression is of the
form subexpression 0 THEN focus on the
subexpression
8
ACT-R Modules The first 2 Seconds 7x338
9
ACT-R Modules The middle 2 Seconds 7x338
10
ACT-R Modules The last 2 Seconds 7x338
11
Learning over 6 Days of Experiment
12
Comments on the ACT-R Model

1. Virtue It actually does the task -- interacts
   with same software as subjects.
2. Virtue The model is not hand crafted but learns
   from instruction (albeit the instructions are a
   little hand-crafted to facilitate parsing).
3. Fact Two parameters were estimated to fit the
   latency data -- the latency scale for retrieval
   and the visual encoding time.
4. Doubt There is an great deal of theoretical
   complexity for a rather simple set of numbers.
5. Resolution We will use brain imaging to test for
   distinct patterns predicted by different modules
   in the model.

13
ACT-R Modules and Buffers
ACT-R
Parse 3x-57
Visual Perception
Fusiform Gyrus
Manual Control
Type x4
Motor Cortex
Production System
Basal Ganglia
Declarative Memory
Retrieve 7512
Prefrontal Cortex
ProblemState
Hold 3x12
Parietal Cortex
Control State
Unwinding Retrieving
Anterior Cingulate
14
Our Modules (all left lateralized) as 100
(5x5x4) Voxel Regions
15
Our Modules (all left lateralized) as 100
(5x5x4) Voxel Regions
16
21.6 Second Structure of fMRI Trial
17
Mapping Module Activity onto the BOLD Response
Activation
18
(No Transcript)
19
(No Transcript)
20
(No Transcript)
21
Prefrontal/Retrieval BA 45/46 (x -40, y 21,
z 21)
22
(No Transcript)
23
Rather directly reflects time because of skipping
steps in equation representation
Parietal/Imaginal BA 39/40 (x -23, y -64, z
34)
24
Rather directly reflects time because of
production rule collapsing
25
c2 measures of Match between Regions and
Modules--small is good (lt130 nonsignificant)
Identical Response Different Peaks
Operation Large Learning Weak
Operation Medium Learning Medium
Operation Medium Learning Medium
Little Response in 0 Operation
26
Observations about fMRI and Modeling

1. While the analysis has been about ACT-R fitting
   the learning of algebra the same methods can be
   used to relate many different information-processi
   ng theories to many tasks.
2. The unifying concept in all cases is that the
   BOLD response in a region reflects time a module
   is engaged. This allows us to map between an
   information-processing model and the BOLD
   response and so to track individual components of
   the model.
3. The same prespecified areas behave as predicted
   in many adult studies.
4. There is no claim one way or another about
   whether the modules are implemented in these
   regions.
5. The critical fact is that we have a measure of
   the activity of specific modules rather than just
   the overall behavior.
6. Challenge Can we take this same model and fit
   it to another experiment.

27
The Second Experiment-- Qin, Sohn, Anderson,
Stenger, Fissel, Goode, Carter (2003)

1. Adults
2. Day 0 Instruction and general practice
3. Days 1 - 5 Computer-based practice
4. Subject types answer by pressing thumb and then
   quickly keying 4 terms.
5. Scanned on Days 1 5.

28
18 Second Structure of fMRI Trial
Blank
Give
Prior
Equation
Period
Answer
?Px?4lt-gt?5
1-3-5-3-4
1.5 Second Scans
29
Instructions for ACT-R

1. To solve an equation, first find the lt-gt, then
   encode the first pair that follows, then shift
   attention to the next pair if there is one, then
   encode the second pair.
2. If this is a simple equation output it otherwise
   process the left side.
3. To process the left side, first find the P.
4. If lt-gt immediately follows then work on the
   operator that precedes the P otherwise first
   encode the pair that follows, then invert the
   operator, and then work on the operator that
   precedes the P.
5. To process the operator that preceded the P,
   first retrieve the transformation associated with
   that operator, then apply the transformation, and
   then output.
6. To output press 1, then output the first, then
   output the next, then output the next, and then
   output the next

Knowledge of inverses (2-3, 4-5) and
transformation rules for getting rid of 2,3,4,
5 prefixes.
30
ACT-R Modules 2 P 3 4 lt-gt 2 5 Encoding
31
ACT-R Modules 2 P 3 4 lt-gt 2 5 Transforming
32
ACT-R Modules 2 P 3 4 lt-gt 2 5 Output
33
Learning over 6 Days of Experiment
34
As before, BOLD response tracks response timing
35
As before, large effects of both factors -- weak
response for 0
Prefrontal/Retrieval BA 45/46 (x -40, y 21,
z 21)
36
As before, large effect for complexity, little
for learning
37
Large effect complexity, learning largely
complete by Day 1
Parietal/Imaginal BA 39/40 (x -23, y -64, z
34)
38
Very Weak Response in this Experiment -- yielding
poor signal to noise ratio.
39
c2 measures of Match between Regions and
Modules--small is good (lt90 nonsignificant)
Identical Response Different Peaks
Operation Large Learning Near Zero
Operation Medium Learning Weak
Poor Signal to Noise Weak Day 5 Response
Little Response in 0 Operation
40
(No Transcript)
41
10 Future Directions for ACT-R

1. Increased stress on parameter-free predictions.
2. Increased effort to anchor the module structure
   of ACT-R with brain correlations.
3. Focus on instruction -- starting our models from
   the beginning.
4. Goal of producing a simulated student.
5. Focus on reasoning and metacognitive processing.
6. Continued effort at community support.
7. Greater emphasis on re-use of components/models
8. Including making knowledge basis available to
   community -- for instance, a middle-school math
   module.
9. Finally get concerned with representational
   assumptions.
10. Facilitate exchange of components between
    architectures.

Local Science
Shared Science