Influence Score

Influence scores quantify the impact of individual data points or nodes on a machine learning model's predictions or behavior. Current research focuses on developing efficient algorithms, such as gradient-based methods and those leveraging graph neural networks, to compute these scores for various model architectures, including large language models and diffusion models. This work is significant for improving model transparency, enabling more efficient data sampling and model training, and facilitating the identification of influential factors in complex systems like social networks and autonomous driving datasets. The ability to accurately assess influence has broad implications for debugging, data curation, and understanding model behavior.