Artificial intelligence models are increasingly integrated into various sectors, requiring a deep understanding of their inner workings. MIT researchers have developed “MAIA,” an automated system that interprets neural networks used in AI vision models. MAIA can identify individual components, clean up irrelevant features, and uncover biases in AI systems. The system combines a vision-language model with interpretability tools to autonomously conduct experiments and refine understanding. While MAIA has room for improvement, it shows potential in auditing AI systems for safety and fairness. The system’s development represents a crucial step in advancing interpretability research and ensuring the resilience of AI technologies.
https://news.mit.edu/2024/mit-researchers-advance-automated-interpretability-ai-models-maia-0723