Seminarium MTPR - Learning to Listen Again: Cycles of Understanding in Artificial Intelligence

Date, time: January 21, 13:00 - 14:00 in the Prof. Arkadiusz Piekara Auditorium, Collegium Physicum

"Learning to Listen Again: Cycles of Understanding in Artificial Intelligence"

Abstract: Our sense of hearing constantly delivers vast amounts of information. In everyday life, our brains are remarkably good at separating what matters from background noise – for example, recognizing a familiar voice in a crowded room or noticing a warning sound in traffic. Teaching computers to do something similar, however, remains a major challenge. While modern artificial intelligence systems can recognize sounds, they often struggle in noisy, real-world environments.
This talk presents research inspired by how humans listen and think. Instead of treating sound analysis as a single, one-step process, the work explores Al systems that process information in repeated cycles much like how people listen, interpret, and reconsider what they hear. This idea is borrowed from models of human cognition, where understanding emerges gradually through multiple passes of perception and reasoning.
The approach is demonstrated in two practical applications. The first focuses on monitoring animal welfare in poultry farming by analyzing chicken vocalizations, helping farmers better understand animal conditions without invasive methods. The second application looks at improving speech clarity, where an Al system repeatedly refines its own output to better separate speech from noise.
Together, these examples show how taking inspiration from human cognition can lead to more robust and flexible Al systems. Such approaches have the potential to improve technologies ranging from healthcare and agriculture to communication tools and everyday sound-based applications.