Trustworthy AI: Controllability and Interpretability in Critical Applications
Project Description
Deep learning (DL) has transformed predictive modeling across different domains, enabling unprecedented accuracy and decision-making capabilities. However, many DL models operate as "black boxes," providing accurate predictions without explaining the underlying reasoning. This lack of transparency poses significant challenges in high-stakes applications, where explainability and controllability are as crucial as predictive performance. The rapid emergence of generative and agentic AI, i.e., systems that autonomously generate content or perform complex tasks, further amplifies these challenges, as users must understand, anticipate, and, when necessary, intervene in model decisions to ensure safety, reliability, and trust.
This project addresses these challenges by investigating methods to enhance the explainability and controllability of deep learning models, with a particular focus on generative and deep learning-based agentic AI. Concept-Based Models (CBMs) are explored as a promising approach, enabling users to explain and control predictions through intermediate, human-understandable variables while maintaining predictive performance. To assess their applicability, the investigated methods will be further evaluated in high-stakes domains, such as healthcare, finance, and industrial automation, exploring their relevance in realworld applications.
This project addresses these challenges by investigating methods to enhance the explainability and controllability of deep learning models, with a particular focus on generative and deep learning-based agentic AI. Concept-Based Models (CBMs) are explored as a promising approach, enabling users to explain and control predictions through intermediate, human-understandable variables while maintaining predictive performance. To assess their applicability, the investigated methods will be further evaluated in high-stakes domains, such as healthcare, finance, and industrial automation, exploring their relevance in realworld applications.
Relevance to Liechtenstein
This project focuses on investigating methods to enhance the explainability and controllability of deep learning models, with a particular emphasis on generative and deep learning-based agentic AI.
These aspects are becoming increasingly important for Liechtenstein, as the country advances its digital transformation and key economic sectors, such as healthcare and finance, face rising demands for AI systems that are reliable and transparent. Moreover, by engaging with local stakeholders through workshops, knowledge-sharing events, and applied collaborations, the project will ensure that its findings contribute meaningfully to the regional innovation ecosystem.
These aspects are becoming increasingly important for Liechtenstein, as the country advances its digital transformation and key economic sectors, such as healthcare and finance, face rising demands for AI systems that are reliable and transparent. Moreover, by engaging with local stakeholders through workshops, knowledge-sharing events, and applied collaborations, the project will ensure that its findings contribute meaningfully to the regional innovation ecosystem.
Scientific, Economic and Societal Impact
This project responds to the growing need for AI systems that can be used in a transparent and responsible way, particularly as generative and agentic models become more integrated into professional and decision-support processes. The ability to understand and, when necessary, intervene in AI-driven reasoning is essential not only for maintaining user trust but also for ensuring the safe and reliable operation of systems that impact socially and economically sensitive areas.
Its practical relevance becomes especially evident in high-stakes domains where AI is already in use. In healthcare, for instance, transparent and controllable models are vital for supporting clinical decisions and ensuring that automated suggestions can be interpreted and assessed by medical professionals. In finance, organizations face increasing expectations regarding fairness, explainability, and regulatory compliance, making opaque models difficult to justify in areas such as risk assessment or fraud detection.
Its practical relevance becomes especially evident in high-stakes domains where AI is already in use. In healthcare, for instance, transparent and controllable models are vital for supporting clinical decisions and ensuring that automated suggestions can be interpreted and assessed by medical professionals. In finance, organizations face increasing expectations regarding fairness, explainability, and regulatory compliance, making opaque models difficult to justify in areas such as risk assessment or fraud detection.
