[PPT] - 2 While building and deploying ML models is now an increasingly PowerPoint Presentation

SLIDE 1

Steven Drucker

Microsoft Research

Combining Visualization and Verbalization for Interpretable Machine Learning

TELEGAM

VIS 2019 Vancouver, Canada Arjun Srinivasan

Georgia Tech

Fred Hohman

@fredhohman

Georgia Tech

SLIDE 2

2

SLIDE 3

While building and deploying ML models is now an increasingly common practice, interpreting models is not.

3

SLIDE 4

4

GAMUT

GAMUT: A Design Probe to Understand How Data Scientists Understand Machine Learning Models. Fred Hohman, Andrew Head, Rich Caruana, Robert DeLine, Steven Drucker. CHI, 2019.

Operationalize

Interpretability in design probe

GAMs

Use generalized additive models

Investigation

Of emerging practice of interpretability w/ industry practitioners

SLIDE 5

Interactive analytics

Visualization

Explanations

Show model context Rely on user interpretation

5

📋

SLIDE 6

Visualization

Explanations

Verbalization

Explanations

Direct and concise Less cognitive load

6

📋 📞

No training needed Interactive analytics Show model context Rely on user interpretation

SLIDE 7

Visualization

Explanations

Verbalization

Explanations

📋 📞

+

7

SLIDE 8

Automatically generate natural language statements,

r verbalizations, to complement explanatory

visualizations for machine learning models.

TELEGAM

Visualization

Explanations

Verbalization

Explanations

📋 📞

+

8

SLIDE 9

Demo

9

SLIDE 10

SLIDE 11

SLIDE 12

Visualize each feature’s global impact on model, grouped by verbalization

SLIDE 13

SLIDE 14

Interactively highlight verbalization in context of the visualization

SLIDE 15

SLIDE 16

Adjust verbalization explanation resolution

SLIDE 17

SLIDE 18

Comparative verbalization

f two prediction

visualizations

SLIDE 19

SLIDE 20

Explanation Resolution

11

SLIDE 21

Predictions vary potentially due to some features contributing differently from both instances.

Explanation Resolution

11

SLIDE 22

Predictions vary potentially due to some features contributing differently from both instances. Predictions vary potentially due to 9 features contributing differently from both instances.

Explanation Resolution

11

SLIDE 23

Predictions vary potentially due to some features contributing differently from both instances. Predictions vary potentially due to 9 features contributing differently from both instances. Predictions 126,024 and 312,129 vary potentially due to 9 features (i.e., 25%) contributing differently from both instances.

Explanation Resolution

11

SLIDE 24

Predictions vary potentially due to some features contributing differently from both instances. Predictions vary potentially due to 9 features contributing differently from both instances. Predictions 126,024 and 312,129 vary potentially due to 9 features (i.e., 25%) contributing differently from both instances.

Explanation Resolution

Brief Detailed

11

SLIDE 25

Verbalization Types

12

Model features

TELEGAM Future Work

Instance features Instance comparison Dataset context Uncertainty …

SLIDE 26

Takeaways

13

SLIDE 27

Takeaways

Visualization + verbalization are complementary

13

📋 + 📞

Combining explanation mediums for the best of both worlds

SLIDE 28

Takeaways

Visualization + verbalization are complementary Use interaction for generation & presentation

13

📋 + 📞

Let users decide resolution, balancing simplicity and completeness Combining explanation mediums for the best of both worlds

💼 + ⚡

SLIDE 29

Steven Drucker

Microsoft Research

Arjun Srinivasan

Georgia Tech

We thank the GT Vis Lab and the anonymous reviewers for their constructive feedback. Funded by a NASA PhD Fellowship.

Fred Hohman

@fredhohman Georgia Tech

Combining Visualization and Verbalization for Interpretable Machine Learning

bit.ly/telegam-vis

Thanks!

TELEGAM

📞

Demo

📙

Paper

🎦

Video

💼

Code

📻

Slides