A Computational Exploration of Problem-Solving Strategies ... · #1029679 and the Georgia Tech...

1
Computational Model We have constructed a working computational model that solves items from the block design task, using techniques from Artificial Intelligence. The system includes modules for perception, attention, memory, and motor action. Each module has parameters that can be changed to represent different individual characteristics. Provided below is an example of the guided search strategy and a 1x1 visual memory size for each block: Direct attention at the target design, and store a mental image of a particular block face and its relative location in the target design. Perform a visual search of the block bank for a matching block If found, grab the matching block. Otherwise, grab a random block and rotate to match stored image. Place the block it in its appropriate location in the construction area. Visual Working Memory Size Mental Image Size = 1x1 Mental Image Size = 3x3 Search Strategy Guided Search Random Search Visual Similarity Threshold Similarity = 0.95 Similarity = 0.65 A Computational Exploration of Problem-Solving Strategies and Gaze Behaviors on the Block Design Task Maithilee Kunda, Mohamed El Banani, and James M. Rehg Computational Perception Lab School of Interactive Computing, Georgia Institute of Technology Maithilee Kunda Vanderbilt University Email: [email protected] Phone: 615-875-8469 Contact 1. Farran, E. K., Jarrold, C., & Gathercole, S. E. (2001). Block design performance in the Williams syndrome phenotype: A problem with mental imagery?. Journal of Child Psychology and Psychiatry, 42(06), 719-728. 2. Weiss, L. G., Saklofske, D. H., Prifitera, A., & Holdnack, J. A. (2006). WISC-IV advanced clinical interpretation. Academic Press 3. Shah, A., & Frith, U. (1993). Why do autistic individuals show superior performance on the block design task?. Journal of Child Psychology and Psychiatry, 34(8), 1351-1364. 4. Kunda M., McGreggor K., & Goel A. K. (2013). A computational model for solving problems from the ravens progressive matrices intelligence test using iconic visual representations. Cognitive Systems Research, 22, 47-66. References The block design task is considered to be an important test of nonverbal reasoning and often represents a peak of ability for people with autism. Our goal is to understand how individual variability on this task is related to fundamental cognitive and neural mechanisms. By building computational models, we can experimentally determine how specific neurocognitive differences lead to different patterns of behavior. Abstract We conducted three experiments in a simulated environment to compare the effects of three different parameters: The Visual Memory Size was varied to allow the model to accurately store a mental image of either a single block, or a 3x3 array of block faces. The Visual Search Strategy was set to either be a random search or a guided search for a specific block face. The Visual Similarity Threshold [4] was set to 0.95 or 0.65. The resulting gaze transition made by the model are presented in Fig. 4. Qualitative differences in the errors produced by the model at lower visual similarity thresholds are shown in Fig. 3. What’s the Block Design Task ? We have shown how slight changes in parameters of a simplified model of human problem solving can result in clear changes in the pattern of observed gaze transitions. We believe that through the development of more complex models, that better match theories of human problem solving, we can shed light on the requirements of the task, as well as the relationship between specific cognitive mechanisms and the resulting behavior. In future work, we will collect data from human participants to analyze their gaze transition graphs and error patterns, and compare them to our computational model to gain more insight into the variability of strategy and behavior across the participant sample. Conclusion and Future Work The Block Design Task is a standardized intelligence test that is commonly used to measure nonverbal reasoning. In a block design task, a person has to reconstruct a design pattern using two-color blocks. Performance on the block design task is sensitive to atypical neuropsychological development, with individuals on the autism spectrum often showing superior performance in the form of faster solution times [1,2]. As a result, the block design task is commonly used in clinical, scientific, and educational settings [3]. Experiment and Results Figure 2. The Computational Model Architecture Figure 1. Overhead view of a participant reconstructing a 4x4 block design task Figure 4. Gaze Transitions made by our computational model. (The independent variable for each comparison is shown to the right. Results show mean +/ standard deviation over 1000 trials) We would like to thank Mika Munch, Yongkoo Kang, Emeke Nkadi, and Richard Stauffer for all their help and insights on this project. This research was supported by a National Science Foundation (NSF) Expedition Award #1029679 and the Georgia Tech Undergraduate Research Opportunities Program (UROP) . Acknowledgments Figure 3. Example Errors for a similarity threshold of 0.64 (left), 0.50 (center), with the target design (right)

Transcript of A Computational Exploration of Problem-Solving Strategies ... · #1029679 and the Georgia Tech...

Page 1: A Computational Exploration of Problem-Solving Strategies ... · #1029679 and the Georgia Tech Undergraduate Research Opportunities Program (UROP) . Acknowledgments Figure 3.Example

Computational ModelWe have constructed a working computational model that solves items from theblock design task, using techniques from Artificial Intelligence. The systemincludes modules for perception, attention, memory, and motor action. Eachmodule has parameters that can be changed to represent different individualcharacteristics.

Provided below is an example of the guided search strategy and a 1x1 visualmemory size for each block:

• Direct attention at the target design, and store a mental image of a particularblock face and its relative location in the target design.

• Perform a visual search of the block bank for a matching block• If found, grab the matching block. Otherwise, grab a random block and rotate

to match stored image.• Place the block it in its appropriate location in the construction area.

Visu

al W

orki

ng M

emor

ySi

ze

Mental Image Size = 1x1 Mental Image Size = 3x3

Sear

ch S

trat

egy

Guided Search Random Search

Visu

al S

imila

rity

Thre

shol

d

Similarity = 0.95 Similarity = 0.65

A Computational Exploration of Problem-Solving Strategies and Gaze Behaviors on the Block Design TaskMaithilee Kunda, Mohamed El Banani, and James M. Rehg

Computational Perception LabSchool of Interactive Computing, Georgia Institute of Technology

Maithilee KundaVanderbilt UniversityEmail: [email protected]: 615-875-8469

Contact1. Farran, E. K., Jarrold, C., & Gathercole, S. E. (2001). Block design performance in the Williams syndrome phenotype: A

problem with mental imagery?. Journal of Child Psychology and Psychiatry, 42(06), 719-728.2. Weiss, L. G., Saklofske, D. H., Prifitera, A., & Holdnack, J. A. (2006). WISC-IV advanced clinical interpretation. Academic Press3. Shah, A., & Frith, U. (1993). Why do autistic individuals show superior performance on the block design task?. Journal of

Child Psychology and Psychiatry, 34(8), 1351-1364.4. Kunda M., McGreggor K., & Goel A. K. (2013). A computational model for solving problems from the ravens progressive

matrices intelligence test using iconic visual representations. Cognitive Systems Research, 22, 47-66.

References

The block design task is considered to be an important testof nonverbal reasoning and often represents a peak ofability for people with autism. Our goal is to understand howindividual variability on this task is related to fundamentalcognitive and neural mechanisms.By building computational models, we can experimentallydetermine how specific neurocognitive differences lead todifferent patterns of behavior.

Abstract

We conducted three experiments in a simulated environment to compare theeffects of three different parameters:• The Visual Memory Size was varied to allow the model to accurately store a

mental image of either a single block, or a 3x3 array of block faces.• The Visual Search Strategy was set to either be a random search or a guided

search for a specific block face.• The Visual Similarity Threshold [4] was set to 0.95 or 0.65.

The resulting gaze transition made by the model are presented in Fig. 4.Qualitative differences in the errors produced by the model at lower visualsimilarity thresholds are shown in Fig. 3.

What’s the Block Design Task ?

We have shown how slight changes in parameters of a simplified model ofhuman problem solving can result in clear changes in the pattern of observedgaze transitions. We believe that through the development of more complexmodels, that better match theories of human problem solving, we can shed lighton the requirements of the task, as well as the relationship between specificcognitive mechanisms and the resulting behavior.

In future work, we will collect data from human participants to analyze their gazetransition graphs and error patterns, and compare them to our computationalmodel to gain more insight into the variability of strategy and behavior across theparticipant sample.

Conclusion and Future Work

The Block Design Task is a standardized intelligence test that is commonly usedto measure nonverbal reasoning. In a block design task, a person has toreconstruct a design pattern using two-color blocks. Performance on the blockdesign task is sensitive to atypical neuropsychological development, withindividuals on the autism spectrum often showing superior performance in theform of faster solution times [1,2]. As a result, the block design task is commonlyused in clinical, scientific, and educational settings [3].

Experiment and Results

Figure 2. The Computational Model Architecture

Figure 1. Overhead view of a participant reconstructing a 4x4 block design task

Figure 4. Gaze Transitions made by our computational model. (The independent variable for each comparison is shown to the right.

Results show mean +/ standard deviation over 1000 trials)

We would like to thank Mika Munch, Yongkoo Kang, Emeke Nkadi, and Richard Stauffer for all their help and insights on this project. This research was supported by a National Science Foundation (NSF) Expedition Award #1029679 and the Georgia Tech Undergraduate Research Opportunities Program (UROP) .

Acknowledgments

Figure 3. Example Errors for a similarity threshold of 0.64 (left), 0.50 (center), with the target design (right)