Interpretability for Generative AI | Kim Montgomery, Navdeep Gill - H2O GenAI Day Atlanta 2024

This session at H2O GenAI Day Atlanta 2024, hosted by Kim Montgomery and Navdeep Gill, focused on the importance of interpretability for Generative AI, particularly Large Language Models (LLMs). The speakers introduced a hands-on approach to understanding and applying interpretability and model risk management techniques in the context of Generative AI. They emphasized the complexity of interpreting unstructured data generated by LLMs, compared to traditional models with structured inputs and outputs.

To address these challenges, they discussed various methods and tools designed to improve the interpretability and reliability of LLMs, including:

➡️ Benchmarking LLMs against specific criteria like accuracy, fairness, and privacy to evaluate their overall performance.
➡️ Utilizing Chain of Verification to systematically verify LLM outputs by generating and answering verification questions based on initial responses.
➡️ Implementing Retrieval Augmented Generation (RAG) for better accuracy checking by comparing model outputs with specific document chunks.
➡️ Exploring influence functions and adversarial testing to assess how changes in input affect LLM outputs and to ensure robust model responses under various conditions.
➡️ Discussing practical applications of these techniques within H2O.ai's platform, such as the Eval Studio for running evaluations and leaderboards for Generative AI applications. Participants were encouraged to engage with the training environment provided by H2O.ai to experiment with these interpretability tools and strategies firsthand.

This interactive workshop highlighted the evolving field of LLM interpretability, offering insights into current best practices and future directions for making Generative AI more understandable, reliable, and trustworthy.

#H2OGenAIDay #AIResponsibility #MachineLearning #GenerativeAI #AIgovernance #EthicalAI