Training Data Influence

Training data influence research investigates how individual training examples impact the performance and behavior of machine learning models, particularly focusing on identifying and mitigating the effects of noisy, biased, or otherwise problematic data. Current research employs various techniques, including influence functions and gradient-based methods, across diverse model architectures such as large language models (LLMs), diffusion models, and deep learning models for image classification. Understanding and addressing training data influence is crucial for improving model robustness, fairness, and explainability, leading to more reliable and trustworthy AI systems across numerous applications.