This project promotes the development of new methods that make artificial intelligence systems more reliable, more data efficient, and easier to correct. Modern systems for language, images, and scientific data often require enormous training sets and computing resources, can become unstable during training, and are difficult to update when information must be removed for privacy or safety reasons. These limitations can hinder scientific discovery and make advanced computing less accessible. This project addresses these challenges by learning how the structure of data shapes the behavior of modern learning systems, with the goal of reducing computational cost, improving reliability, and supporting safer curation of learnt models. The project will also strengthen the future computing workforce through undergraduate and graduate research training, course-based projects, open software and educational materials, and hands-on outreach for school students and teachers on data, algorithms, and responsible artificial intelligence. The research studies how individual training examples shape the local geometry of the loss function in modern machine learning. It has three connected aims. First, it will characterize and improve optimization stability in deep neural networks, including modern predictive and generative models, by developing diagnostics and training methods based on curvature alignment across data. Second, it will design small data summaries and synthetic training sets that preserve the structure of the full learning problem, thereby reducing data and computational cost while maintaining performance. Third, it will develop efficient methods for removing the influence of selected training examples with minimal damage to the rest of the model. The project will evaluate these ideas on image, language, continual learning, and modern text and image generation benchmarks, and will release benchmarks and instructional modules to support reproducible research and educa