Speaker: Patricia Alonso de Apellaniz
Abstract:
This talk traces a research journey from signal processing and audio-based machine learning to generative modeling, survival analysis, and trustworthy AI. It begins by exploring how early work on voice activity detection and audiovisual generation, along with industry experience, shaped an applied research approach. The main part of the talk presents research on deep generative models for healthcare, focusing on survival analysis, synthetic tabular data generation and validation, and federated learning under scarce, heterogeneous, and privacy-sensitive data. It also examines the transition towards real-world deployment through European research projects and the broadening of these methods through research stays at the University of Copenhagen and the Institute of Ceramics and Glass of the Spanish National Research Council (ICV-CSIC), including their transfer to battery diagnostics. Finally, the talk introduces current research directions in interpretable deep learning with Kolmogorov–Arnold Networks, causal inference, and conversational interfaces to make predictive and explanatory models more accessible. Across these topics, a recurring question emerges: how can models be made not only accurate, but also understandable and useful in practice?