Influence Function

Influence functions are a powerful tool for understanding how individual training data points affect a machine learning model's predictions, aiming to improve model interpretability and robustness. Current research focuses on improving the computational efficiency of influence function estimation, particularly for large language models and other complex architectures, often employing techniques like low-rank approximations and gradient-based methods. This work addresses challenges like inaccurate approximations and the limitations of influence functions in non-convex settings, with applications ranging from data selection and anomaly detection to model debugging and fairness analysis. The development of more accurate and efficient influence function methods holds significant promise for enhancing the transparency and reliability of machine learning systems.