Jun.-Prof. Dr. Kira Maag, Department Visual Analytics, HHU

“From Performance to Trust: Making Deep Neural Networks Safe and Robust”

Deep neural networks have achieved remarkable performance in visual perception tasks and are increasingly integrated into real-world systems. In applications such as automated driving, however, high benchmark accuracy alone is not sufficient: perception systems must operate reliably under changing environmental conditions, rare events, and potential external manipulation. In this talk, I will use semantic scene understanding as a guiding example to discuss how deep neural networks can be made safer, more robust, and more reliable. This includes analyzing how models behave under distribution shifts, how uncertainty can be quantified to assess prediction quality, and how adversarial perturbations affect decision boundaries and model stability. The goal is to contribute to a more structured and principle-based approach to trustworthy AI.

Prof. Dr. Kevin Tang, Department English Language and Linguistics, HHU

“Mind the Gap! Bridging Human Communication and AI Systems for Social Good”

Human communication is fundamentally rooted in alignment -- the subconscious synchronization of language and speech that facilitates mutual understanding and collaborative success. However, in high-stakes environments such as healthcare and social services, this alignment often breaks down due to systemic demographic and linguistic gaps between providers and the communities they serve. In real-life clinical settings, while African American English (AAE) clinicians naturally align with AAE patients to build rapport, a severe demographic bias in the clinician population makes such matching a rare luxury.

In this keynote, I first introduce how linguistic alignment predicts human collaborative outcomes, drawing on studies in which people communicate while collaborating on a task. Second, I examine the potentials and challenges in harnessing this phenonmenon in Human and Machine communication. I share findings in the context of cancer screening using virtual health assistants, linguistic alignment can significantly lead to increase trust and willingness to be screened. However, I contrast this potential with technical challenges: contemporary Automatic Speech Recognition (ASR) systems systematically underperform for speakers of marginalized dialects, such as AAE and Newcastle English, creating a barrier to these very interventions. Finally, I discuss methodological solutions for "Minding the Gap": from leveraging oral histories for diverse data harvesting to developing precision annotation tools and forced-alignment techniques for under-resourced languages. By transitioning from a "one-size-fits-all" model to a linguistically informed AI, we can build technology that does not just process language, but actively fosters human connection and social equity.