2025 NVIDIA GTC: Crack the AI Black Box: Practical Techniques for Explainable AI
0 likes•30 views
2025 NVIDIA GTC: Crack the AI Black Box: Practical Techniques for Explainable AI Artificial Intelligence often operates in ways that are challenging to interpret, creating a gap in trust and transparency. Explainable AI (XAI) bridges this gap by providing strategies to demystify complex models, enabling stakeholders to understand how decisions are made. We'll explore foundational XAI concepts and provide practical methods to bring interpretability into developing and deploying AI systems, ensuring better decision-making and accountability. You'll learn actionable techniques for explaining AI behavior, from feature attributions and decision-path analyses to scenario-based insights. Through a live demonstration, you'll see how to apply these methods to real-world problems, enabling you to diagnose, debug, and optimize your models effectively. In the end, you'll have a clear roadmap for integrating XAI practices into your workflows to build trust and confidence in AI-powered solutions. Key Takeaways: Understand how AI systems make decisions, critical for trust and adoption as they grow more complex In 2025, explainable AI (XAI) is set to be a hot topic as industries increasingly demand transparency from AI systems. You'll gain practical techniques to make AI transparent and interpretable, breaking down decision paths and uncovering key insights into model behavior Featuring live demos of explainability methods applied to real-world scenarios, you'll leave with actionable strategies to debug, optimize, and build confidence in your AI solutions
![感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！
ibaotu.com
Resources
[CLICK HERE] for All Materials and Demos in this Session
DigitalOcean Bare Metal H200 Availability
https://www.digitalocean.com/blog/now-available-bare-metal-nvidia-hgx-h200-gpus
Continue the Conversation - DigitalOcean Discord
https://discord.com/invite/digitalocean
● Captum:
○ GitHub - https://github.com/pytorch/captum
○ Tutorials - https://captum.ai/tutorials/
● PyTorch:
○ GitHub - https://github.com/pytorch/pytorch
○ Tutorials - https://pytorch.org/tutorials/index.html](https://image.slidesharecdn.com/cracktheaiblackboxpracticaltechniquesforexplainableai-250328195718-e22f78b2/85/2025-NVIDIA-GTC-Crack-the-AI-Black-Box-Practical-Techniques-for-Explainable-AI-28-320.jpg)
2025 NVIDIA GTC: Crack the AI Black Box: Practical Techniques for Explainable AI
- 1. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com David vonThenen @davidvonthenen Crack the AI Black Box PracticalTechniques for Explainable AI
- 2. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com ● Are you Human or an AI? ● I want 5 Kubernetes ● Virtual Machines are Real ● Cloudy, cloudy, cloudy… ● There is storage for that! David vonThenen @davidvonthenen
- 3. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 3 Agenda ● What is Explainable AI? ● Understanding Data Inconsistencies ● Dataset Observability and Diagnostics ○ Demo: NLP ○ Demo: Vision ● Adversarial Attacks for Good… & Bad ● Let's Continue the Discussion!
- 5. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 5 Flawed Data ● AI/ML Only As Good As the Data ○ What If the Data is Flawed? ● Real-World Examples: ○ Recruiter AI + Male Skewed ○ Offensive AI Chatbot ○ Court Case Hallucinations ○ Home Purchasing Spree ○ Scan Sitting Up or Down ○ Many, Many, Many More
- 6. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 6 Explainable AI Why Do We Care? ● Trust and Transparency ● Debugging + Improvement ● Compliance and Ethics Key Goals: ● Interpretability ● Accountability ● Fairness + Bias Detection
- 8. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 8 Data Inconsistencies Matter ● Impacts AI "Decision Making" ○ Annotation Errors ○ Data Bias ○ Distribution Drift ○ Adversarial Input ● Accidental, Unintended Consequences
- 10. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 10 Data Bias Image Attribution Princeton University Researcher - Tool helps clear biases from computer vision Molly Sharlach on October 1, 2020
- 11. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 11 Data Imbalance Image Attribution Introduction to Balanced and Imbalanced Datasets in Machine Learning Nikolaj Buhl on November 11, 2022
- 12. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 12 Distribution Shifts Image Attribution Dive Into Deep Learning Nikolaj Buhl on November 11, 2022
- 13. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 13 Adversarial Samples Image Attribution: PyTorch Documentation - Adversarial Example Generation
- 15. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 15 What Tools Can I Use? ● Captum - https://github.com/pytorch/captum ● SHAP - https://github.com/shap/shap ● LIME ● ELI5 ● AIX360 ● Many More
- 16. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 16 Let's Take a Look at Captum ● Open Source PyTorch Library ○ Gradients, Saliency Maps, SHAP ○ Layer/Neuron Contributions ○ NLP, Vision ● Detects: ○ Biases ○ Inconsistency ○ Hidden Patterns
- 17. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 17 Captum: Case Study ● Study: Urinary Incontinence ● Captum Revealed Findings: ○ Validated Contributions ○ Discovered 3 Features ● Future Application: ○ Update Surgical Protocols ○ Improved Techniques ○ Post-Op Therapy Case Study Paper An artificial intelligence method for predicting postoperative urinary incontinence based on multiple anatomic parameters of MRI
- 18. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com Demo: Captum + NLP Classifier https://youtu.be/geZNwLzoaT4 https://youtu.be/m0VxUAGhKcY
- 19. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com Demo: Captum + Vision Classifier https://youtu.be/5J2sGIU0RV4
- 20. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com Adversarial Attacks: For Good… and Bad Building Better Models via Intentional Disruption
- 21. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 21 Turning Insights Into Action ● Why Explainable AI? ○ Question Rigid Assumptions ○ Finding Data Flaws ○ Expose Ethical Scenarios ○ Adversarial Testing ● Result ○ Why Exclude Data ○ Fix Problematic Data ○ Underrepresentation ○ Fairness
- 22. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 22 What Else… ● Intentional Adversarial Attacks ○ Besides Finding Holes… ○ Disrupting Classification ■ Vision ■ NLP ● Why? ○ Unauthorized Surveillance ○ Protect Privacy ○ Obfuscation
- 23. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 23 Adversarial Strategies Ideas/Concepts in NLP to Disrupt Be Creative!! ● Encoding/Formatting ● Homophones and Phonetics ● Code Switching ● Low-Resource Languages ○ Navajo - "Code Talkers" ● Adversarial Spelling ● Polysemy/Multiple Meanings ● Speaking in Metaphors Video Credit: Darmok, under 8 minutes (Star Trek Abridged) Source: Star Trek: The Next Generator, Episode 102 - Darmok
- 28. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com Resources [CLICK HERE] for All Materials and Demos in this Session DigitalOcean Bare Metal H200 Availability https://www.digitalocean.com/blog/now-available-bare-metal-nvidia-hgx-h200-gpus Continue the Conversation - DigitalOcean Discord https://discord.com/invite/digitalocean ● Captum: ○ GitHub - https://github.com/pytorch/captum ○ Tutorials - https://captum.ai/tutorials/ ● PyTorch: ○ GitHub - https://github.com/pytorch/pytorch ○ Tutorials - https://pytorch.org/tutorials/index.html
- 29. 感谢您下载包图网平台上提供的PPT作品，为了您和包图网以及原创作者的利益，请勿复制、传播、销售，否则将承担法律责任！包图网将对作品进行维权，按照传播下载次数进行十倍的索取赔偿！ ibaotu.com 29 Thank You! Senior AI/ML Engineer @davidvonthenen David vonThenen