SlideShare uma empresa Scribd logo
1 de 21
ActiVis
Visual Exploration of Industry-Scale
Deep Neural Network Models
Minsuk Kahng Pierre Andrews Aditya Kalro Polo Chau
Georgia Tech Georgia TechFacebook Facebook
*
Deep Learning is Powerful
2Image credit:http://www.nvidia.com/object/drive-px.html, https://venturebeat.com/2016/04/14/, https://finance.yahoo.com/news/
Understanding Deep Learning is Challenging
3
Cat
Dog
Cat
Dog
INPUT OUTPUTMODEL
Image credit: https://www.kaggle.com/c/dogs-vs-cats/
Incorrect
Cat
Dog
Visualizing Instance Activation
4
Which neurons are highly activated for a given input?
INPUT OUTPUTMODEL
Cat
Dog 78%
22%
[Yosinski et al., 2015; Harley, 2015; Karpathy, 2016; Liu et al., 2017]
tags: #mycat, #cute
date: 10/1/2017
location: 33.7, 88.4
Practical Challenges in Industry
5
DIVERSE INPUT TYPESCOMPLEX MODELS LARGE DATASETS
image, text, numerical,
categorical, …
many nodes
in graph-structure
1 billion+
instances
1. 2. 3.
Develop ActiVis for Facebook-scale models and dataGOAL:
Enjoying nice weather with kiki❤️
Understanding ActiVis Users’ Needs
Participatory design sessions over 11 months, with
15+ Facebook researchers, engineers, & data scientists
6
Complex Model Architectures
UNDERSTANDING ACTIVIS USERS’ NEEDS (1/3)
7
Numerous deep & wide models are used.
input
output
Separate architecture from activation detailsGOAL:
Huge Datasets with Diverse Features
UNDERSTANDING ACTIVIS USERS’ NEEDS (2/3)
1 billion+
instances
8
1,000+ multi-type features
(e.g., image, text, numerical, categorical)
tags: #mycat, #cute
date: 10/1/2017
location: 33.7, 88.4
Enjoying nice weather with kiki❤️
Use multiple approaches for scalability
GOAL:
Make use of diverse features
Two Key Analytics Patterns
UNDERSTANDING ACTIVIS USERS’ NEEDS (3/3)
How model responds to
individual instances?
(This instance highly activates
neurons #2, 5, 11.)
How model behaves at higher-level
categorization (e.g., by topic)?
(A subset of instances about “sports” highly
activates neurons #3, 7, 11.)
Useful for debugging Useful for large datasets
SUBSET-LEVELINSTANCE-LEVEL
9
[Kahng et al., 2016; Krause et al., 2016]
Complementary
[Kulesza et al., 2015; Amershi et al., 2015]
Support both analyticsGOAL:
ActiVis Design Goals (Recap)
10
Two analytics patterns
Complex models
Huge datasets
1.
2.
3. Unified analysis for instances & subsets
Model overview as entry point
Multiple approaches for scalability
ActiVis
Visualizing activation of industry-scale deep neural nets
Deployed on Facebook ML Platform
11
12
demo:
“Where is Phoenix located?” → location
“What is the diameter of a golf ball?” → numeric
Exploring text classification results
ActiVis Demo
Demo
13
• Unified analysis for instances & subsets
• Model architecture to activation details
• Scaling to industry-scale data & models
14
ActiVis Key Ideas (Demo Recap)
User-guided Instance Sampling
15
SCALING TO LARGE DATA & MODELS (1/2)
e.g., “What does VAST mean?”
should be in ABBR class.
Users either want representative samples or maintain “test cases”.
Selective Precomputation for Important Nodes
16
SCALING TO LARGE DATA & MODELS (2/2)
Often only a few nodes are helpful and model developers know them.
precomputed
Deployed on FBLearner
17
Model developers add 3 API calls to enable ActiVis.
Facebook ML platform, used by 25% of their engineers
Click to launch
ActiVis
Case Studies
3 Facebook participants.
All work with text classification models.
Each session 60 minutes long.
18
Spot1. -checking models with “test cases”
“Where is … located?” → location
Graph architecture view as entry point2.
Debugging hints from activation patterns3.
19
Key Observations from Case Studies
• Discover interesting subsets interactively
• Support input-dependent models (e.g., RNN)
• Provide direct guidance for performance improvement
20
Future Research Directions
ActiVis
Visual Exploration of Industry-Scale
Deep Neural Network Models
✓ Deployed on Facebook’s ML platform
✓ Support subset-level analysis
✓ Model architecture to activation details
We thank Facebook Applied Machine Learning Group, especially
Yangqing Jia, Andrew Tulloch, Liang Xiong, and Zhao Tan and
NSF Graduate Research Fellowship Program.
Pierre Andrews
Aditya Kalro
Polo Chau
Georgia Tech
Facebook
Facebook
Minsuk Kahng
Georgia Tech PhD student
http://minsuk.com

Mais conteúdo relacionado

Mais procurados

MLFlow: Platform for Complete Machine Learning Lifecycle
MLFlow: Platform for Complete Machine Learning Lifecycle MLFlow: Platform for Complete Machine Learning Lifecycle
MLFlow: Platform for Complete Machine Learning Lifecycle
Databricks
 
Monitoring AI with AI
Monitoring AI with AIMonitoring AI with AI
Monitoring AI with AI
Stepan Pushkarev
 
Machine Learning Interpretability - Mateusz Dymczyk - H2O AI World London 2018
Machine Learning Interpretability - Mateusz Dymczyk - H2O AI World London 2018Machine Learning Interpretability - Mateusz Dymczyk - H2O AI World London 2018
Machine Learning Interpretability - Mateusz Dymczyk - H2O AI World London 2018
Sri Ambati
 

Mais procurados (15)

MLFlow: Platform for Complete Machine Learning Lifecycle
MLFlow: Platform for Complete Machine Learning Lifecycle MLFlow: Platform for Complete Machine Learning Lifecycle
MLFlow: Platform for Complete Machine Learning Lifecycle
 
CD4ML and the challenges of testing and quality in ML systems
CD4ML and the challenges of testing and quality in ML systemsCD4ML and the challenges of testing and quality in ML systems
CD4ML and the challenges of testing and quality in ML systems
 
Deep Learning with CNTK
Deep Learning with CNTKDeep Learning with CNTK
Deep Learning with CNTK
 
Eclipse DemoCamp Budapest 2016 November: Best of EclipseCon Europe 2016
Eclipse DemoCamp Budapest 2016 November: Best of EclipseCon Europe 2016Eclipse DemoCamp Budapest 2016 November: Best of EclipseCon Europe 2016
Eclipse DemoCamp Budapest 2016 November: Best of EclipseCon Europe 2016
 
Monitoring AI with AI
Monitoring AI with AIMonitoring AI with AI
Monitoring AI with AI
 
Introduction to Distributed Computing Engines for Data Processing - Simone Ro...
Introduction to Distributed Computing Engines for Data Processing - Simone Ro...Introduction to Distributed Computing Engines for Data Processing - Simone Ro...
Introduction to Distributed Computing Engines for Data Processing - Simone Ro...
 
第1回バイオインフォマティクスデータ可視化セミナー@Riken
第1回バイオインフォマティクスデータ可視化セミナー@Riken第1回バイオインフォマティクスデータ可視化セミナー@Riken
第1回バイオインフォマティクスデータ可視化セミナー@Riken
 
Dependency inversion using ports and adapters
Dependency inversion using ports and adaptersDependency inversion using ports and adapters
Dependency inversion using ports and adapters
 
Whats new in Eclipse Indigo ? (@DemoCamp Grenoble 2011)
Whats new in Eclipse Indigo ? (@DemoCamp Grenoble 2011)Whats new in Eclipse Indigo ? (@DemoCamp Grenoble 2011)
Whats new in Eclipse Indigo ? (@DemoCamp Grenoble 2011)
 
Machine Learning Interpretability - Mateusz Dymczyk - H2O AI World London 2018
Machine Learning Interpretability - Mateusz Dymczyk - H2O AI World London 2018Machine Learning Interpretability - Mateusz Dymczyk - H2O AI World London 2018
Machine Learning Interpretability - Mateusz Dymczyk - H2O AI World London 2018
 
AISF19 - Building Scalable, Kubernetes-Native ML/AI Pipelines with TFX, KubeF...
AISF19 - Building Scalable, Kubernetes-Native ML/AI Pipelines with TFX, KubeF...AISF19 - Building Scalable, Kubernetes-Native ML/AI Pipelines with TFX, KubeF...
AISF19 - Building Scalable, Kubernetes-Native ML/AI Pipelines with TFX, KubeF...
 
[Research] azure ml anatomy of a machine learning service - Sharat Chikkerur
[Research] azure ml  anatomy of a machine learning service - Sharat Chikkerur[Research] azure ml  anatomy of a machine learning service - Sharat Chikkerur
[Research] azure ml anatomy of a machine learning service - Sharat Chikkerur
 
Ai use cases
Ai use casesAi use cases
Ai use cases
 
Developing Digital Twins
Developing Digital TwinsDeveloping Digital Twins
Developing Digital Twins
 
Build 2019 Recap
Build 2019 RecapBuild 2019 Recap
Build 2019 Recap
 

Semelhante a ActiVis: Visual Exploration of Industry-Scale Deep Neural Network Models

Semelhante a ActiVis: Visual Exploration of Industry-Scale Deep Neural Network Models (20)

TechEvent 2019: Artificial Intelligence in Dev & Ops; Martin Luckow - Trivadis
TechEvent 2019: Artificial Intelligence in Dev & Ops; Martin Luckow - TrivadisTechEvent 2019: Artificial Intelligence in Dev & Ops; Martin Luckow - Trivadis
TechEvent 2019: Artificial Intelligence in Dev & Ops; Martin Luckow - Trivadis
 
Ai open powermeetupmarch25th
Ai open powermeetupmarch25thAi open powermeetupmarch25th
Ai open powermeetupmarch25th
 
Ultra Fast Deep Learning in Hybrid Cloud Using Intel Analytics Zoo & Alluxio
Ultra Fast Deep Learning in Hybrid Cloud Using Intel Analytics Zoo & AlluxioUltra Fast Deep Learning in Hybrid Cloud Using Intel Analytics Zoo & Alluxio
Ultra Fast Deep Learning in Hybrid Cloud Using Intel Analytics Zoo & Alluxio
 
Whole Chain Traceability, pulling a Kobayashi Maru.
Whole Chain Traceability, pulling a Kobayashi Maru. Whole Chain Traceability, pulling a Kobayashi Maru.
Whole Chain Traceability, pulling a Kobayashi Maru.
 
Data science lab enabling flexibility
Data science lab   enabling flexibilityData science lab   enabling flexibility
Data science lab enabling flexibility
 
Understanding the New World of Cognitive Computing
Understanding the New World of Cognitive ComputingUnderstanding the New World of Cognitive Computing
Understanding the New World of Cognitive Computing
 
Whole Chain Traceability Consortium
Whole Chain Traceability ConsortiumWhole Chain Traceability Consortium
Whole Chain Traceability Consortium
 
research Paper face recognition attendance system
research Paper face recognition attendance systemresearch Paper face recognition attendance system
research Paper face recognition attendance system
 
Ai open powermeetupmarch25th
Ai open powermeetupmarch25thAi open powermeetupmarch25th
Ai open powermeetupmarch25th
 
Ai open powermeetupmarch25th
Ai open powermeetupmarch25thAi open powermeetupmarch25th
Ai open powermeetupmarch25th
 
Software Analytics: Towards Software Mining that Matters (2014)
Software Analytics:Towards Software Mining that Matters (2014)Software Analytics:Towards Software Mining that Matters (2014)
Software Analytics: Towards Software Mining that Matters (2014)
 
Efficient Data Labelling for Ocular Imaging
Efficient Data Labelling for Ocular ImagingEfficient Data Labelling for Ocular Imaging
Efficient Data Labelling for Ocular Imaging
 
Applied AI Workshop - Presentation - Connect Day GDL
Applied AI Workshop - Presentation - Connect Day GDLApplied AI Workshop - Presentation - Connect Day GDL
Applied AI Workshop - Presentation - Connect Day GDL
 
Technovision
TechnovisionTechnovision
Technovision
 
Feature store: Solving anti-patterns in ML-systems
Feature store: Solving anti-patterns in ML-systemsFeature store: Solving anti-patterns in ML-systems
Feature store: Solving anti-patterns in ML-systems
 
Visualization for Software Analytics
Visualization for Software AnalyticsVisualization for Software Analytics
Visualization for Software Analytics
 
Big Data LDN 2018: HOW AUTOMATION CAN ACCELERATE THE DELIVERY OF MACHINE LEAR...
Big Data LDN 2018: HOW AUTOMATION CAN ACCELERATE THE DELIVERY OF MACHINE LEAR...Big Data LDN 2018: HOW AUTOMATION CAN ACCELERATE THE DELIVERY OF MACHINE LEAR...
Big Data LDN 2018: HOW AUTOMATION CAN ACCELERATE THE DELIVERY OF MACHINE LEAR...
 
chapter 5.docx
chapter 5.docxchapter 5.docx
chapter 5.docx
 
chapter 5.pdf
chapter 5.pdfchapter 5.pdf
chapter 5.pdf
 
Emerging engineering issues for building large scale AI systems By Srinivas P...
Emerging engineering issues for building large scale AI systems By Srinivas P...Emerging engineering issues for building large scale AI systems By Srinivas P...
Emerging engineering issues for building large scale AI systems By Srinivas P...
 

Último

TECUNIQUE: Success Stories: IT Service provider
TECUNIQUE: Success Stories: IT Service providerTECUNIQUE: Success Stories: IT Service provider
TECUNIQUE: Success Stories: IT Service provider
mohitmore19
 
CALL ON ➥8923113531 🔝Call Girls Kakori Lucknow best sexual service Online ☂️
CALL ON ➥8923113531 🔝Call Girls Kakori Lucknow best sexual service Online  ☂️CALL ON ➥8923113531 🔝Call Girls Kakori Lucknow best sexual service Online  ☂️
CALL ON ➥8923113531 🔝Call Girls Kakori Lucknow best sexual service Online ☂️
anilsa9823
 
CALL ON ➥8923113531 🔝Call Girls Badshah Nagar Lucknow best Female service
CALL ON ➥8923113531 🔝Call Girls Badshah Nagar Lucknow best Female serviceCALL ON ➥8923113531 🔝Call Girls Badshah Nagar Lucknow best Female service
CALL ON ➥8923113531 🔝Call Girls Badshah Nagar Lucknow best Female service
anilsa9823
 
CHEAP Call Girls in Pushp Vihar (-DELHI )🔝 9953056974🔝(=)/CALL GIRLS SERVICE
CHEAP Call Girls in Pushp Vihar (-DELHI )🔝 9953056974🔝(=)/CALL GIRLS SERVICECHEAP Call Girls in Pushp Vihar (-DELHI )🔝 9953056974🔝(=)/CALL GIRLS SERVICE
CHEAP Call Girls in Pushp Vihar (-DELHI )🔝 9953056974🔝(=)/CALL GIRLS SERVICE
9953056974 Low Rate Call Girls In Saket, Delhi NCR
 

Último (20)

SyndBuddy AI 2k Review 2024: Revolutionizing Content Syndication with AI
SyndBuddy AI 2k Review 2024: Revolutionizing Content Syndication with AISyndBuddy AI 2k Review 2024: Revolutionizing Content Syndication with AI
SyndBuddy AI 2k Review 2024: Revolutionizing Content Syndication with AI
 
Unveiling the Tech Salsa of LAMs with Janus in Real-Time Applications
Unveiling the Tech Salsa of LAMs with Janus in Real-Time ApplicationsUnveiling the Tech Salsa of LAMs with Janus in Real-Time Applications
Unveiling the Tech Salsa of LAMs with Janus in Real-Time Applications
 
HR Software Buyers Guide in 2024 - HRSoftware.com
HR Software Buyers Guide in 2024 - HRSoftware.comHR Software Buyers Guide in 2024 - HRSoftware.com
HR Software Buyers Guide in 2024 - HRSoftware.com
 
W01_panagenda_Navigating-the-Future-with-The-Hitchhikers-Guide-to-Notes-and-D...
W01_panagenda_Navigating-the-Future-with-The-Hitchhikers-Guide-to-Notes-and-D...W01_panagenda_Navigating-the-Future-with-The-Hitchhikers-Guide-to-Notes-and-D...
W01_panagenda_Navigating-the-Future-with-The-Hitchhikers-Guide-to-Notes-and-D...
 
How To Use Server-Side Rendering with Nuxt.js
How To Use Server-Side Rendering with Nuxt.jsHow To Use Server-Side Rendering with Nuxt.js
How To Use Server-Side Rendering with Nuxt.js
 
The Real-World Challenges of Medical Device Cybersecurity- Mitigating Vulnera...
The Real-World Challenges of Medical Device Cybersecurity- Mitigating Vulnera...The Real-World Challenges of Medical Device Cybersecurity- Mitigating Vulnera...
The Real-World Challenges of Medical Device Cybersecurity- Mitigating Vulnera...
 
TECUNIQUE: Success Stories: IT Service provider
TECUNIQUE: Success Stories: IT Service providerTECUNIQUE: Success Stories: IT Service provider
TECUNIQUE: Success Stories: IT Service provider
 
Hand gesture recognition PROJECT PPT.pptx
Hand gesture recognition PROJECT PPT.pptxHand gesture recognition PROJECT PPT.pptx
Hand gesture recognition PROJECT PPT.pptx
 
5 Signs You Need a Fashion PLM Software.pdf
5 Signs You Need a Fashion PLM Software.pdf5 Signs You Need a Fashion PLM Software.pdf
5 Signs You Need a Fashion PLM Software.pdf
 
CALL ON ➥8923113531 🔝Call Girls Kakori Lucknow best sexual service Online ☂️
CALL ON ➥8923113531 🔝Call Girls Kakori Lucknow best sexual service Online  ☂️CALL ON ➥8923113531 🔝Call Girls Kakori Lucknow best sexual service Online  ☂️
CALL ON ➥8923113531 🔝Call Girls Kakori Lucknow best sexual service Online ☂️
 
Optimizing AI for immediate response in Smart CCTV
Optimizing AI for immediate response in Smart CCTVOptimizing AI for immediate response in Smart CCTV
Optimizing AI for immediate response in Smart CCTV
 
Try MyIntelliAccount Cloud Accounting Software As A Service Solution Risk Fre...
Try MyIntelliAccount Cloud Accounting Software As A Service Solution Risk Fre...Try MyIntelliAccount Cloud Accounting Software As A Service Solution Risk Fre...
Try MyIntelliAccount Cloud Accounting Software As A Service Solution Risk Fre...
 
CALL ON ➥8923113531 🔝Call Girls Badshah Nagar Lucknow best Female service
CALL ON ➥8923113531 🔝Call Girls Badshah Nagar Lucknow best Female serviceCALL ON ➥8923113531 🔝Call Girls Badshah Nagar Lucknow best Female service
CALL ON ➥8923113531 🔝Call Girls Badshah Nagar Lucknow best Female service
 
CHEAP Call Girls in Pushp Vihar (-DELHI )🔝 9953056974🔝(=)/CALL GIRLS SERVICE
CHEAP Call Girls in Pushp Vihar (-DELHI )🔝 9953056974🔝(=)/CALL GIRLS SERVICECHEAP Call Girls in Pushp Vihar (-DELHI )🔝 9953056974🔝(=)/CALL GIRLS SERVICE
CHEAP Call Girls in Pushp Vihar (-DELHI )🔝 9953056974🔝(=)/CALL GIRLS SERVICE
 
Reassessing the Bedrock of Clinical Function Models: An Examination of Large ...
Reassessing the Bedrock of Clinical Function Models: An Examination of Large ...Reassessing the Bedrock of Clinical Function Models: An Examination of Large ...
Reassessing the Bedrock of Clinical Function Models: An Examination of Large ...
 
Tech Tuesday-Harness the Power of Effective Resource Planning with OnePlan’s ...
Tech Tuesday-Harness the Power of Effective Resource Planning with OnePlan’s ...Tech Tuesday-Harness the Power of Effective Resource Planning with OnePlan’s ...
Tech Tuesday-Harness the Power of Effective Resource Planning with OnePlan’s ...
 
Vip Call Girls Noida ➡️ Delhi ➡️ 9999965857 No Advance 24HRS Live
Vip Call Girls Noida ➡️ Delhi ➡️ 9999965857 No Advance 24HRS LiveVip Call Girls Noida ➡️ Delhi ➡️ 9999965857 No Advance 24HRS Live
Vip Call Girls Noida ➡️ Delhi ➡️ 9999965857 No Advance 24HRS Live
 
Unlocking the Future of AI Agents with Large Language Models
Unlocking the Future of AI Agents with Large Language ModelsUnlocking the Future of AI Agents with Large Language Models
Unlocking the Future of AI Agents with Large Language Models
 
Software Quality Assurance Interview Questions
Software Quality Assurance Interview QuestionsSoftware Quality Assurance Interview Questions
Software Quality Assurance Interview Questions
 
The Ultimate Test Automation Guide_ Best Practices and Tips.pdf
The Ultimate Test Automation Guide_ Best Practices and Tips.pdfThe Ultimate Test Automation Guide_ Best Practices and Tips.pdf
The Ultimate Test Automation Guide_ Best Practices and Tips.pdf
 

ActiVis: Visual Exploration of Industry-Scale Deep Neural Network Models

  • 1. ActiVis Visual Exploration of Industry-Scale Deep Neural Network Models Minsuk Kahng Pierre Andrews Aditya Kalro Polo Chau Georgia Tech Georgia TechFacebook Facebook *
  • 2. Deep Learning is Powerful 2Image credit:http://www.nvidia.com/object/drive-px.html, https://venturebeat.com/2016/04/14/, https://finance.yahoo.com/news/
  • 3. Understanding Deep Learning is Challenging 3 Cat Dog Cat Dog INPUT OUTPUTMODEL Image credit: https://www.kaggle.com/c/dogs-vs-cats/ Incorrect
  • 4. Cat Dog Visualizing Instance Activation 4 Which neurons are highly activated for a given input? INPUT OUTPUTMODEL Cat Dog 78% 22% [Yosinski et al., 2015; Harley, 2015; Karpathy, 2016; Liu et al., 2017]
  • 5. tags: #mycat, #cute date: 10/1/2017 location: 33.7, 88.4 Practical Challenges in Industry 5 DIVERSE INPUT TYPESCOMPLEX MODELS LARGE DATASETS image, text, numerical, categorical, … many nodes in graph-structure 1 billion+ instances 1. 2. 3. Develop ActiVis for Facebook-scale models and dataGOAL: Enjoying nice weather with kiki❤️
  • 6. Understanding ActiVis Users’ Needs Participatory design sessions over 11 months, with 15+ Facebook researchers, engineers, & data scientists 6
  • 7. Complex Model Architectures UNDERSTANDING ACTIVIS USERS’ NEEDS (1/3) 7 Numerous deep & wide models are used. input output Separate architecture from activation detailsGOAL:
  • 8. Huge Datasets with Diverse Features UNDERSTANDING ACTIVIS USERS’ NEEDS (2/3) 1 billion+ instances 8 1,000+ multi-type features (e.g., image, text, numerical, categorical) tags: #mycat, #cute date: 10/1/2017 location: 33.7, 88.4 Enjoying nice weather with kiki❤️ Use multiple approaches for scalability GOAL: Make use of diverse features
  • 9. Two Key Analytics Patterns UNDERSTANDING ACTIVIS USERS’ NEEDS (3/3) How model responds to individual instances? (This instance highly activates neurons #2, 5, 11.) How model behaves at higher-level categorization (e.g., by topic)? (A subset of instances about “sports” highly activates neurons #3, 7, 11.) Useful for debugging Useful for large datasets SUBSET-LEVELINSTANCE-LEVEL 9 [Kahng et al., 2016; Krause et al., 2016] Complementary [Kulesza et al., 2015; Amershi et al., 2015] Support both analyticsGOAL:
  • 10. ActiVis Design Goals (Recap) 10 Two analytics patterns Complex models Huge datasets 1. 2. 3. Unified analysis for instances & subsets Model overview as entry point Multiple approaches for scalability
  • 11. ActiVis Visualizing activation of industry-scale deep neural nets Deployed on Facebook ML Platform 11
  • 12. 12 demo: “Where is Phoenix located?” → location “What is the diameter of a golf ball?” → numeric Exploring text classification results ActiVis Demo
  • 14. • Unified analysis for instances & subsets • Model architecture to activation details • Scaling to industry-scale data & models 14 ActiVis Key Ideas (Demo Recap)
  • 15. User-guided Instance Sampling 15 SCALING TO LARGE DATA & MODELS (1/2) e.g., “What does VAST mean?” should be in ABBR class. Users either want representative samples or maintain “test cases”.
  • 16. Selective Precomputation for Important Nodes 16 SCALING TO LARGE DATA & MODELS (2/2) Often only a few nodes are helpful and model developers know them. precomputed
  • 17. Deployed on FBLearner 17 Model developers add 3 API calls to enable ActiVis. Facebook ML platform, used by 25% of their engineers Click to launch ActiVis
  • 18. Case Studies 3 Facebook participants. All work with text classification models. Each session 60 minutes long. 18
  • 19. Spot1. -checking models with “test cases” “Where is … located?” → location Graph architecture view as entry point2. Debugging hints from activation patterns3. 19 Key Observations from Case Studies
  • 20. • Discover interesting subsets interactively • Support input-dependent models (e.g., RNN) • Provide direct guidance for performance improvement 20 Future Research Directions
  • 21. ActiVis Visual Exploration of Industry-Scale Deep Neural Network Models ✓ Deployed on Facebook’s ML platform ✓ Support subset-level analysis ✓ Model architecture to activation details We thank Facebook Applied Machine Learning Group, especially Yangqing Jia, Andrew Tulloch, Liang Xiong, and Zhao Tan and NSF Graduate Research Fellowship Program. Pierre Andrews Aditya Kalro Polo Chau Georgia Tech Facebook Facebook Minsuk Kahng Georgia Tech PhD student http://minsuk.com