Dernière mise à jour :
Building AI systems that jointly learn multiple tasks (e.g., speech recognition + translation), with governance needed for complexity and auditability.
Multi-task architectures that share representations and weights across related objectives, improving sample efficiency but increasing coupling and opacity. Governance challenges include ensuring each task meets its performance and fairness requirements, managing complex deployment pipelines, and maintaining explainability. Documentation must clearly delineate how shared components affect each task’s outputs and how updates propagate across tasks.
Exemple concret :
A global-customer-support AI uses joint modeling for classification (“issue category”) and sentiment analysis. When retraining, the team separately evaluates fairness and accuracy for both tasks, logs each task’s metrics, and reviews any cross-impact (e.g., sentiment biases affecting classification). Governance ensures that improvements in one task do not degrade the other.
