11-07, 10:55–11:35 (US/Eastern), Music Box
Large Language Models (LLMs) offer new avenues for explaining and debugging machine learning models through natural language interfaces. This talk explores how LLMs can interpret both interpretable models, such as Generalized Additive Models (GAMs), and complex black-box models using post-hoc methods. By analyzing modular components of interpretable models, LLMs can provide insights without exceeding context window limitations. We also demonstrate how LLMs leverage their extensive prior knowledge to detect anomalies and suggest potential issues in models. Attendees will learn practical techniques for using LLMs to enhance model transparency and trust in AI systems.
You can find the slides and the code here: https://github.com/avilog/shap2llm/blob/main/examples/XAI%20LLMs.pptx
Understanding and debugging complex machine learning models is becoming increasingly difficult, especially in high-stakes domains like finance and healthcare. This talk presents how LLMs can bridge this gap by improving explainability and transparency:
1) Interpreting Modular Models: Learn how LLMs can describe, interpret, and debug models decomposed into modular components, focusing on GAMs that represent outcomes as sums of univariate and pairwise functions.
2) Chain-of-Thought Reasoning: See how LLMs apply sequential reasoning to analyze each component separately, allowing for scalable insights within limited context windows.
3) Leveraging Prior Knowledge: Discover how LLMs use their extensive domain knowledge to detect surprises and anomalies in models, highlighting potential issues in data collection, model fitting, or specification.
4) Explaining Black-Box Models: Explore how combining LLMs with post-hoc explanation methods like SHAP enables the interpretation of complex, opaque models.
5) Risks: Understand key challenges such as LLM hallucinations, cost concerns, and privacy issues and practical solutions to mitigate these risks.
6) Practical Demonstrations: Gain hands-on insights with examples using Python libraries and LLM APIs.
By the end of the session, you'll have the tools to harness LLMs for explaining and debugging both self-interpretable and black-box machine-learning models, while balancing transparency with the risks involved.
No previous knowledge expected
Avi is an AI tech lead at Citi Innovation Lab.
He holds a master's in computer science with a thesis on interpretable machine learning.