Here are my software architecture keynote slides on “Distributed Trust Architecture – The New Reality of ML-based Systems.” It’s based on the latest scientific papers and projects from CSIRO’s Data61 Thanks to all the co-authors!

Abstract: Rarely are ML-based systems in the real world about training a model over some centralised datasets and then integrating the model into a deployed system. The architecture complexity of modern ML-based systems is multifaceted. Data and compute often have to be federated or at the edge due to data ownership, security, privacy, regulatory and ethical constraints. Data, models, code and configurations give rise to intermingled dependencies, not just within a single organisation but in an ecosystem. Model-as-a-service (including “foundation models” and their adaptation) and “continuous everything” introduces new uncertainties. This talk outlines the challenges and solutions in software architecture to help address these challenges via distributed trust, federated ML-based systems, design patterns and other architecture-level assurance mechanisms.