![8
Our ShapeShifter Attack: Stop Sign ! Person
Spotlighted in new DARPA GARD program [PKDD’18; with Intel]
Real Stop Sign
Printed Adversarial
Stop Sign](https://image.slidesharecdn.com/human-centered-ai-polo-chau-190423181122/85/Human-Centered-AI-Scalable-Interactive-Tools-for-Interpretation-and-Attribution-8-320.jpg)
![8
Our ShapeShifter Attack: Stop Sign ! Person
Spotlighted in new DARPA GARD program [PKDD’18; with Intel]](https://image.slidesharecdn.com/human-centered-ai-polo-chau-190423181122/85/Human-Centered-AI-Scalable-Interactive-Tools-for-Interpretation-and-Attribution-9-320.jpg)
![SHIELD: Fast Practical Defense for
Deep Learning via JPEG Compression
[KDD’18 Audience Appreciation Award runner up; with intel]
9](https://image.slidesharecdn.com/human-centered-ai-polo-chau-190423181122/85/Human-Centered-AI-Scalable-Interactive-Tools-for-Interpretation-and-Attribution-10-320.jpg)
![How to make tools easy-to-use for
various users?
Challenge 3
EXPERTS NOVICES
[Hohman, Kahng, Pienta, Chau, TVCG, 2018]
PRACTITIONERS](https://image.slidesharecdn.com/human-centered-ai-polo-chau-190423181122/85/Human-Centered-AI-Scalable-Interactive-Tools-for-Interpretation-and-Attribution-24-320.jpg)
![ActiVis
Scaling Visualization to
Industry-scale Models and Data
Deployed by[Kahng, et al. IEEE VIS’17]](https://image.slidesharecdn.com/human-centered-ai-polo-chau-190423181122/85/Human-Centered-AI-Scalable-Interactive-Tools-for-Interpretation-and-Attribution-26-320.jpg)
![ML Cube
Interactive Model Comparison
Deployed by[Kahng, et al. HILDA’16]](https://image.slidesharecdn.com/human-centered-ai-polo-chau-190423181122/85/Human-Centered-AI-Scalable-Interactive-Tools-for-Interpretation-and-Attribution-40-320.jpg)
![GAN Lab
Interactive understanding of
complex deep learning models
PAIR | People + AI Research Initiative
[Kahng, et al. IEEE VIS’18]](https://image.slidesharecdn.com/human-centered-ai-polo-chau-190423181122/85/Human-Centered-AI-Scalable-Interactive-Tools-for-Interpretation-and-Attribution-48-320.jpg)
![Generative Adversarial Networks (GANs)
46
“the most interesting idea in the last 10 years in ML”
- Yann LeCun
Face images generated by BEGAN [Berthelot et al., 2017]](https://image.slidesharecdn.com/human-centered-ai-polo-chau-190423181122/85/Human-Centered-AI-Scalable-Interactive-Tools-for-Interpretation-and-Attribution-51-320.jpg)
Human-Centered AI: Scalable, Interactive Tools for Interpretation and Attribution
- 1. S C A L A B L E I N T E R A C T I V E T O O L S I N T E R P R E T A T I O N & A T T R I B U T I O N HUMAN-CENTERED AI Polo Chau Georgia Tech Associate Professor, Computational Science & Engineering ML Area Leader, College of Computing Associate Director, MS Analytics for
- 2. AI HI+ HUMAN INTELLIGENCE ARTIFICIAL INTELLIGENCE Scalable interactive tools to make sense of complex large-scale datasets and models Polo Club of Data Science
- 3. Human-Centered AI Cyber Security Social Good & HealthLarge Graph Mining & Visualization Polo Club of Data Science poloclub.github.io Adversarial ML
- 4. Why should we make AI more human-centered? Accessible & Interpretable for people who build and use machine learning
- 5. ?
- 6. AI models often used as black-box
- 7. New York Times, 2018
- 8. 8 Our ShapeShifter Attack: Stop Sign ! Person Spotlighted in new DARPA GARD program [PKDD’18; with Intel] Real Stop Sign Printed Adversarial Stop Sign
- 9. 8 Our ShapeShifter Attack: Stop Sign ! Person Spotlighted in new DARPA GARD program [PKDD’18; with Intel]
- 10. SHIELD: Fast Practical Defense for Deep Learning via JPEG Compression [KDD’18 Audience Appreciation Award runner up; with intel] 9
- 11. We do NOT know why AI attacks and defenses work Actually nobody does (e.g., which neurons are under attack?)
- 12. Image used in xkcd.com Kim, CVPR’18 tutorial Image credit:
- 13. Practitioners very interested in why and how AI works Figure from Washington Post
- 16. RESEARCH PROBLEM How can we design scalable, interactive, usable interfaces for people to understand complex, large-scale ML systems?
- 17. OUR KEY IDEA Scalable Interactive visualization as a medium for connecting users with ML models 15
- 18. Why interactive visualization? Machine learning aims to find patterns from data. Visualization amplifies human cognition to find patterns.
- 19. Why interactive visualization? By interacting with visualization, users can incrementally make sense of AI models.
- 20. How to visualize complex models? 18 Challenge 1 GoogLeNet Neural Machine Translation
- 21. How to scale up to large datasets? Challenge 2
- 22. How to scale up to large datasets? Challenge 2
- 23. How to scale up to large datasets? Challenge 2 ?
- 24. How to make tools easy-to-use for various users? Challenge 3 EXPERTS NOVICES [Hohman, Kahng, Pienta, Chau, TVCG, 2018] PRACTITIONERS
- 25. Human-Centered AI by Visual Analytics Our Research GAN Lab ActiVis Visualization for Industry-Scale Models Interactive Learning of Complex Models ML Cube - Model comparison - Activation analysis by subsets - Experimentation with GANs
- 26. ActiVis Scaling Visualization to Industry-scale Models and Data Deployed by[Kahng, et al. IEEE VIS’17]
- 27. why practitioners want visualization? Facebook data scientists need visualization tools to interpret complex models
- 28. Those down totags: #mycat, #cute date: 10/1/2017 location: 33.7, 88.4 Practical Design Challenges 24 DIVERSE FEATURESDEEP/WIDE MODELS LARGE DATASETS image, text, numerical, categorical, … 1,000+ operations/layers 1 billion+ instances Enjoying nice weather with kiki❤
- 29. UNDERSTANDING USERS’ NEEDS Participatory design sessions with 15+ researchers, engineers & data scientists at Facebook over 11 months 25
- 30. Visualizing activation of industry-scale deep neural nets, deployed by Facebook 26 ActiVis
- 31. ActiVis Research Challenges 1. Many model parameters to visualize 2. Many data instances to analyze 3. Intensive computation for deployment 27 Visualization for industry-scale deep models
- 32. Location How to visualize many model parameters? 28 Challenge #1 INPUT OUTPUTMODEL Person Location 81% 8% Where is Mercedes-Benz Stadium located? Number 11% many layers
- 33. Location How to visualize many model parameters? 28 Challenge #1 INPUT OUTPUTMODEL Person Location 81% 8% Where is Mercedes-Benz Stadium located? Number 11% particularly useful many layers Observation: No need to show everything
- 34. Model Overview to Activation Details 29 ActiVis Key Ideas #1
- 35. How to analyze many data instances? 30 Challenge #2 SUBSET-LEVELINSTANCE-LEVEL Complementary Useful for debugging Useful for large datasets Observation: Two Analytics Patterns How model behaves at higher-level categorization (e.g., by topic)? How model responds to individual instances?
- 36. Unified Analysis for Instances & Subsets ActiVis Key Ideas (2)
- 37. Scaling Up ActiVis for Facebook 1. User-guided Instance Sampling 2. Selective Pre-computation of Layers 3. Matrix Computation for Billion-Scale Instances 32 ActiVis Key Ideas (3)
- 38. Deployed on FBLearner 33 used by >25% of engineering team Facebook’s ML platform
- 39. Human-Centered AI by Visual Analytics Our Research GAN Lab ActiVis Visualization for Industry-Scale Models Interactive Learning of Complex Models ML Cube - Model comparison - Activation analysis by subsets - Experimentation with GANs
- 40. ML Cube Interactive Model Comparison Deployed by[Kahng, et al. HILDA’16]
- 41. Challenge: Model Selection Which model to use? Baseline Model New Model 89.5% 90.1%overall accuracy Age 20-39 Age 13-19 92.0% 97.0% 87.0% 69.0%
- 42. Comparison by Subsets with Data Cube 37 country age gender age gender country Model A accuracy Model B accuracy * * * 89.5% 90.1% 13-19 * * 92.0% 97.0% 20-39 * * 87.0% 69.0% * F * 89.6% 89.9% 13-19 F * 91.0% 93.0% 13-19 F USA 91.1% 94.0% * M USA 75.5% 74.0% 20-39 M Canada 87.2% 73.7% How to scale to very large number of possible subsets?
- 43. 38 Visual Model Comparison via Interactive Data Cube Exploration MLCube
- 44. 39
- 45. Impact Deployed on Facebook’s ML Platform Influenced Google’s new open-source system, TensorFlow Model Analysis 40
- 46. Research Contributions from ActiVis & MLCube How can we scale visualization to industry-scale models and data? 1.Exploration from overview to details 2.Drilling down into specific parts of data 3.Combination of scalable & interactive methods (Under review — discover interesting subsets automatically: FairVis: Visual Analytics for Discovering Intersectional Bias in Machine Learning)
- 47. Human-Centered AI by Visual Analytics Our Research GAN Lab ActiVis Visualization for Industry-Scale Models Interactive Learning of Complex Models ML Cube - Model comparison - Activation analysis by subsets - Experimentation with GANs
- 48. GAN Lab Interactive understanding of complex deep learning models PAIR | People + AI Research Initiative [Kahng, et al. IEEE VIS’18]
- 49. Visualization for ML Education
- 50. Modern deep models are complex 45
- 51. Generative Adversarial Networks (GANs) 46 “the most interesting idea in the last 10 years in ML” - Yann LeCun Face images generated by BEGAN [Berthelot et al., 2017]
- 52. Generative Adversarial Networks (GANs) Hard to understand and train even for experts 47 Discriminator Generator
- 53. Why GANs are hard? A GAN uses two competing neural networks 48 Discriminator spots fake Police spots fake bills Generator synthesizes outputs Counterfeiter makes fake bills
- 54. Why GANs are hard? A GAN uses two competing neural networks 48 Discriminator spots fake Police spots fake bills Generator synthesizes outputs Counterfeiter makes fake bills
- 55. Why GANs are hard? A GAN uses two competing neural networks 48 Discriminator spots fake Police spots fake bills Generator synthesizes outputs Counterfeiter makes fake bills
- 56. Why GANs are hard? A GAN uses two competing neural networks 48 Discriminator spots fake Police spots fake bills Generator synthesizes outputs Counterfeiter makes fake bills
- 57. Why GANs are hard? A GAN uses two competing neural networks 48 Discriminator spots fake Police spots fake bills Generator synthesizes outputs Counterfeiter makes fake bills
- 58. Why GANs are hard? A GAN uses two competing neural networks 48 Discriminator spots fake Police spots fake bills Generator synthesizes outputs Counterfeiter makes fake bills
- 59. GAN Lab Research Challenges 1. Conceptual understanding of GANs 2. Interactive model training 3. Easily accessible for students Can we design an interactive tool for GANs?
- 61. What type of data to visualize? 50 Discriminator (Police) Generator (Counterfeiter)
- 62. What type of data to visualize? 2D distribution, instead of high-dimensional images 50 Discriminator (Police) Generator (Counterfeiter)
- 63. What type of data to visualize? 2D distribution, instead of high-dimensional images 51 Discriminator (Police) Generator (Counterfeiter)
- 64. What type of data to visualize? 2D distribution, instead of high-dimensional images 51 Discriminator (Police) Generator (Counterfeiter) 1. To focus on GAN’s main concepts 2. To easily visualize data distribution Why 2D data points?
- 66. Discriminator (Police) How to visually explain the generator? 53 Generator (Counterfeiter)
- 67. Discriminator (Police) How to visually explain the generator? 53 Generator (Counterfeiter)
- 68. How to visually explain the generator? 54 Generator (Counterfeiter)
- 69. How to visually explain the generator? random 54 Generator (Counterfeiter)
- 70. How to visually explain the generator? map an input point into a new position random 54 Generator (Counterfeiter)
- 71. How to visually explain the generator? map an input point into a new position random ? 54 Generator (Counterfeiter)
- 72. How to visually explain the generator? map an input point into a new position random ? 54 Generator (Counterfeiter)
- 73. How to visually explain the generator? map an input point into a new position random ? 54 Generator (Counterfeiter)
- 74. How to visually explain the generator? map an input point into a new position random Manifold ? 54 Generator (Counterfeiter)
- 75. How to visually explain the generator? map an input point into a new position random 54 Generator (Counterfeiter)
- 76. How to visualize the discriminator? 55 Generator (Counterfeiter) Discriminator (Police)
- 77. How to visualize the discriminator? 55 Generator (Counterfeiter) Discriminator (Police)
- 78. How to visualize the discriminator? 56
- 79. How to visualize the discriminator? 2D heatmap, to represent binary classification 56 Data points in this region are likely real. Data points are likely fake.
- 80. VER. 0.5 57 realdata fakedata generator discriminator + + +
- 81. VER. 0.5 58 Hard to develop mental models for GANs
- 84. 1. Building mental models for GANs 2. Locating hyperparameters 3. Tracking data flow How GAN Lab helps?
- 85. 61
- 86. GAN Lab broadens education access 62 Conventional Deep Learning Visualization in JavaScript in Python with GPU Model Training Visualization $$$
- 87. Everything done in browser, powered by TensorFlow.js GAN Lab broadens education access 63 Accelerated by WebGL in JavaScript Visualization also in JavaScript Model Training
- 88. GAN Lab is Live! 64 30K visitors, 135 countries 1.9K Likes 800+ Retweets Try at bit.ly/gan-lab
- 89. Research Contributions from GAN Lab Can we design tools for non-experts to understand complex deep models? 1.Visualization of overall structure & components 2.Interactive experimentation of training process 3.Accessible approach using browsers
- 90. Visual Analytics in Deep Learning: An Interrogative Survey for the Next Frontiers Fred Hohman, Minsuk Kahng, Robert Pienta, Polo Chau TVCG 2018 !66
- 91. !67 Some Takeaways 1. Most tools aimed at expert users 2. Instance-based analysis 3. Inherently interdisciplinary 4. Lacks actionability 5. Evaluation is hard 6. State-of-the-art models not robust bit.ly/va-dl-survey
- 92. How can we make AI more human-centered? Accessible & Interpretable for people who build and use machine learning
- 93. How can we make AI more human-centered? Accessible & Interpretable for people who build and use machine learning Through the design of visualization tools that are scalable, interactive & usable, we can help users learn and interpret large-scale complex ML systems.
- 94. Thanks! Polo Chau Georgia Tech HUMAN-CENTERED AI S C A L A B L E I N T E R A C T I V E T O O L S I N T E R P R E T A T I O N & A T T R I B U T I O N for