July 25, 2024 – Researchers at the Computer Science and Artificial Intelligence Laboratory (CSAIL) of the Massachusetts Institute of Technology (MIT) have developed a system called “MAIA,” a multimodal automated interpretability agent that utilizes visual language models to automatically carry out various neural network interpretability tasks.
MAIA, which stands for Multimodal Automated Interpretability Agent, leverages visual language models to perform a range of neural network interpretability tasks automatically. Additionally, it comes equipped with tools for experimentation on other artificial intelligence systems.
According to Tamar Rott Shaham, a post-doctoral researcher at MIT CSAIL and co-author of the research paper, “Our goal is to create an AI researcher capable of conducting interpretability experiments independently.” Current automated interpretability methods, Shaham explained, typically involve annotating or visualizing data in a one-time process.
However, MAIA goes beyond this by generating hypotheses, designing experiments to test them, and refining its understanding through iterative analysis. By integrating pre-trained visual language models with a library of interpretability tools, MAIA’s multimodal approach allows it to assemble and run targeted experiments on specific models, responding to user inquiries and refining its methods until it can provide comprehensive answers.
This automated agent has proven capable of accomplishing three crucial tasks: labeling the internal components of visual models and describing the visual concepts that activate them, cleaning up image classifiers by eliminating irrelevant features to enhance their robustness in new situations, and identifying hidden biases in AI systems to uncover potential fairness issues in their outputs.
MAIA’s ability to generate hypotheses, test them through experiments, and improve its understanding through iterative analysis offers valuable insights into the inner workings of AI models. This not only aids in comprehending how AI models operate but also in exploring their safety and potential biases.
