Interpretable AI: A Hands-on Introduction
2h 17mAdvanced2026-12-16
Authors
Soumya Batra
Course details
What is this course about?
Join instructor Soumya Batra as she outlines a practical, skills-focused approach to coding, equipping you with advanced techniques and strategies to peek inside the proverbial "black box" of language models. Through a series of guided exercises and demos, you’ll progress from visualizing basic attention patterns to more advanced scenarios such as implementing cutting-edge methods like sparse autoencoders and steering vectors. This course balances theoretical understanding with hands-on experience, giving you an opportunity to apply these techniques immediately to solve real-world AI development challenges.
Objectives
What will I be able to do by the end of this course?
Differentiate interpretability from explainability in AI, identifying the most valuable scenarios for each technique.
Apply and analyze attention map visualizations and sparse autoencoders to understand model internals.
Implement and leverage steering vectors for practical applications, influencing model behavior.
Train and evaluate your own sparse autoencoders to uncover meaningful concepts within transformer models.
Utilize theoretical and hands-on knowledge to improve AI model transparency and efficacy in real-world scenarios.
Audience
Who is this course for?
Machine learning engineers
AI researchers
Data scientists
Technical professionals working with or developing language models
Join instructor Soumya Batra as she outlines a practical, skills-focused approach to coding, equipping you with advanced techniques and strategies to peek inside the proverbial "black box" of language models. Through a series of guided exercises and demos, you’ll progress from visualizing basic attention patterns to more advanced scenarios such as implementing cutting-edge methods like sparse autoencoders and steering vectors. This course balances theoretical understanding with hands-on experience, giving you an opportunity to apply these techniques immediately to solve real-world AI development challenges.
Objectives
What will I be able to do by the end of this course?
Differentiate interpretability from explainability in AI, identifying the most valuable scenarios for each technique.
Apply and analyze attention map visualizations and sparse autoencoders to understand model internals.
Implement and leverage steering vectors for practical applications, influencing model behavior.
Train and evaluate your own sparse autoencoders to uncover meaningful concepts within transformer models.
Utilize theoretical and hands-on knowledge to improve AI model transparency and efficacy in real-world scenarios.
Audience
Who is this course for?
Machine learning engineers
AI researchers
Data scientists
Technical professionals working with or developing language models
Concepts
Introduction
- Under the hood of interpretable AI
Interpretability - A Tool, Not a Rule
- Should you care about AI interpretability
Visualizing Attention Maps
- Unlocking attention maps - From setup to model loading
- Visualizing attention maps, part 1
- Visualizing attention maps, part 2
The Five Courses of Sparse Autoencoders (SAE)
- The star chef of interpretability - SAE
- Amuse-bouche - Environment setup and model loading
- Appetizer - Data loading training and extracting
- Main course - Training SAE
- Palate cleanser - Verify SAE correctness and health
- Dessert - Identifying interesting features using SAE
- Challenge - Find a hidden concept
- Solution - Find a hidden concept
- Challenge - Train a SAE
Steering Vectors
- Using steering vectors on models
- Demo - Setting up and training steering vectors
- Demo - Fun with steering vectors
- Challenge - Influence model behavior using steering vectors
- Solution - Influence model behavior using steering vectors