From Gradients To Governance: Making PyTorch Lineage-Aware - Kateryna Romashko & Clodagh Walsh

This talk, presented by Kateryna Romashko and Clodagh Walsh from Red Hat, explores the necessity of integrating data lineage within PyTorch to enhance governance in AI systems. The speakers emphasize that as AI operates with regulated and jurisdiction-bound data, it is essential for models to not only learn but also adhere to lineage and policy constraints. They propose a framework where PyTorch's dynamic graphs and autograd system can support metadata related to origin, consent, and policy throughout training and inference. The focus is on creating a lineage-aware PyTorch that fosters trustworthy and auditable AI solutions across various ecosystems, including edge and federated environments.