Download presentation
Presentation is loading. Please wait.
Published byFabiana Fantini Modified over 5 years ago
1
Drag and Track: A Direct Manipulation Interface for Contextualizing Data Instances within a Continuous Parameter Space Daniel Orban, University of Minnesota Daniel F. Keefe, University of Minnesota Ayan Biswas, Los Alamos National Laboratory James Ahrens, Los Alamos National Laboratory David Rogers, Los Alamos National Laboratory
2
Drag and Track: A Direct Manipulation Interface for Contextualizing Data Instances within a Continuous Parameter Space Daniel Orban, University of Minnesota Daniel F. Keefe, University of Minnesota Ayan Biswas, Los Alamos National Laboratory James Ahrens, Los Alamos National Laboratory David Rogers, Los Alamos National Laboratory
3
What would you do with an aluminum alloy?
Pressure
4
What if you were a material scientist?
Pressure
5
A material scientist would want to shoot it.
Pressure
6
Velocity Pressure Temperature What if…? Credit: JPL-Calteck/NASA
Credit: NASA's Goddard Space Flight Center/SDO Photo: Jerry Lara /San Antonio Express-News
7
Running a shock physics experiment.
8
The shock physics input output model.
Velocimetry Profile
9
Run the experiment enough times with different inputs and you get an ensemble.
10
Run the experiment enough times with different inputs and you get an ensemble.
Goals: Understand the high-dimensional parameter space. Explore the gaps in the ensemble. Understand the input from the output.
11
Roadmap Motivation Related Work Drag and Track
Results, Evaluation, and Conclusions
12
To assist with parameter space analysis we use a direct manipulation interface within a dimensionally reduced space. Visual Parameter Space Analysis Direct Manipulation within Dimensionality Reduction Sedlmair et al. 2014, IEEE VIS Cavallo and Demiralp, 2018, ACM CHI Endert et al. 2011, IEEE VIS Input and Output Spaces Navigational Strategies Local-to-Global Global-to-Local Prediction Uncertainty / Sensitivity Understanding dimensionally reduced views Spatial relationships have meaning Underlying mechanisms are hidden from the user
13
Like related work in SciVIS our direct manipulation works in both forward and inverse directions.
Forward Design Inverse Design Design by Dragging, Coffey, et al. 2013, IEEE VIS
14
We know it is critical to visualize prediction uncertainty together with the ensemble.
Torsney-Weir et al. 2011, IEEE VIS Berger et al. 2011, Computer Graphics Forum
15
Roadmap Motivation Related Work Drag and Track
Results, Evaluation, and Conclusions
16
Key Idea: Directly manipulate Virtual Data Instances (VDI) to interactively explore a parameter space. Virtual Data Instance
17
Key Idea: Directly manipulate Virtual Data Instances (VDI) to interactively explore a parameter space. Virtual Data Instance
18
Key Idea: Directly manipulate Virtual Data Instances (VDI) to interactively explore a parameter space. Virtual Data Instance Interactive Callout
19
Key Idea: Directly manipulate Virtual Data Instances (VDI) to interactively explore a parameter space. Virtual Data Instance Interactive Callout
20
Key Idea: Directly manipulate Virtual Data Instances (VDI) to interactively explore a parameter space. Virtual Data Instance Interactive Callout Nearest Neighbor Highlights
21
Drag and Track Interface
Parallel Coordinates Plot for Filtering in Either Space Input Space Output Space
22
Drag and Track Interface
Parallel Coordinates Plot for Filtering in Either Space Input Space Output Space
23
Example Task 1: Explore the parameter space through annotation.
24
Example Task 2: Track parameter changes by dragging the elastic plateau.
25
Example Task 3: Study the sensitivity by moving VDIs across the input and output spaces.
26
What is a Virtual Data Instance?
Parameters Name: Output 137 Value: 3.14 Weight: 0.75 Data Instance P1 P2 P3 ... Pn Name: Pressure Value: 2.718 Name: PCA 1 Value: 3.14 Index: 27 Weight: 0.125 Virtual Data Instance Query Set: P2 P7 ... P23 Predicted Data Instance: P1 P2 P3 ... Pn Nearest Neighbors (index): N1 N2 ... Nk
27
Roadmap Motivation Related Work Drag and Track
Results, Evaluation, and Conclusions
28
Input parameter A qualitatively correlates with the elastic plateau output feature.
29
Domain Expert Evaluation
Domain Experts: 2 Shock Physics Experimentalists 1 Fluid Simulation Expert Time: Each user evaluated Drag and Track for minutes. Methodology: The first 10 minutes involved walking the user through the application. The last minutes was a talk out loud discussion. A series of questions were asked related to each of the features in Drag and Track and whether they contributed to the following goals: Understanding the input and output space Uncertainty and sensitivity Directly manipulating the output to gain insight about the input.
30
Direct manipulation enables sensitivity analysis by changing one value and observing the others.
I think certainly, there is alot there to look at with sensitivity. [I can] move around one input value, all at the same point,... to measure sensitivity of different parameters... As we drag the elastic plateau, [we] would we be able to see range in specific parameters. Was there one particular parameter that moved highly through the path?
31
DR views are still abstract, but the nearest neighbor spread across spaces intuitively shows model uncertainty. You can really tell you the nearest neighbors… It is a little abstract understanding where I am in the PCA space and seeing how that is a real parameter. In terms of how [the input and output space] would be related, it does give me some information about jumping between the two spaces… [The uncertainty] distribution is covered by nearest nearest neighbors.
32
Tracks show model correlation between the input and output spaces.
Input Path (large curve) Output Path (tight curve) I really like drag lines that show up… the clean line in the parameter space shows how it has a tight curve in one place and a large [curve] in the other… That [shows] how well the model correlates..
33
Overall Feedback Great for quickly exploring new data sets.
Opportunity to match experiment output curves with inverse direct manipulation.
34
Summary and Conclusions
Drag & Track uses the concept of "virtual data instances" to enable forward and inverse direct manipulation within dimensionally reduced visualizations of high-dimensional parameter spaces. Several extensions are also introduced, including interactive callouts, track lines, and nearest-neighbor highlighting to convey uncertainty. Our initial application and evaluations with domain scientists (along with experiments on a synthetic dataset described in the paper) suggest Drag & Track can be useful for identifying trends and performing visual parameter sensitivity analysis. We believe the visuals and integrated interaction techniques show promise for helping users contextualize ensembles in continuous parameter spaces across a variety of other domains as well.
35
Thank you! Acknowledgements Shock Physics Feedback
David Walters Richard Sandberg Cindy Bolme Color Map Francesca Samsel Software is available and open source! Cinema Quest: This document was published under LA-UR
37
Key Idea: Directly manipulate Virtual Data Instances (VDI) to interactively explore a parameter space in context.
38
Drag and Track users can both manipulate VDIs and annotate spaces
39
Evaluation Tasks
40
Evaluation Tasks
41
Evaluation Tasks
42
Drag and Track - Contextualizing Data Instances
43
Velocity Pressure Temperature What if…? 8 km/sec
Credit: JPL-Calteck/NASA What if…? Velocity Pressure Temperature 8 km/sec Credit: NASA's Goddard Space Flight Center/SDO Photo: Jerry Lara /San Antonio Express-News
44
Shock Physics Light Gas Gun Typically hydrogen
Projectile speeds up to 8km / sec (¾ the velocity needed to escape the Earth’s gravity) 100 to 200 shots per year
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.