Constraint Guided Deep Learning to Add Semantics to Acoustic Signals
Webinar with Peter Karsmakers (KU Leuven DTAI).
Practical information:
Leertraject
Want to register?
- Register until: 24 Mar 2022
- Price: free
Registration is not necessary, you can simply join the seminar.
Registration is possible to receive a reminder of the seminars in this series on time series.
Georganiseerd door:
Acoustic signals are a rich source of information that have ample perspective to glean useful insights about the monitoring context. Furthermore, information can be collected at a distance from the acoustic sources without making physical contact. Deep learning models are typically used to automatically add semantics to the acoustics. However, such models typically are learned using a massive amount of input-output pairs, ignoring domain knowledge that additionally might be available. When domain knowledge is injected by customizing model architectures and/or constraining the learning mechanism, it is expected that the need for annotated data is relaxed. Furthermore, when the processing platform on which the models will be deployed is known beforehand learning can already take into account constraints that entail thresholds on computational and memory complexity. In this way the model deployment process can be smoothened. After explaining the research context some concrete use-cases will be discussed along with future work directions.
Acoustic signals are a rich source of information that have ample perspective to glean useful insights about the monitoring context. Furthermore, information can be collected at a distance from the acoustic sources without making physical contact. Deep learning models are typically used to automatically add semantics to the acoustics. However, such models typically are learned using a massive amount of input-output pairs, ignoring domain knowledge that additionally might be available. When domain knowledge is injected by customizing model architectures and/or constraining the learning mechanism, it is expected that the need for annotated data is relaxed. Furthermore, when the processing platform on which the models will be deployed is known beforehand learning can already take into account constraints that entail thresholds on computational and memory complexity. In this way the model deployment process can be smoothened. After explaining the research context some concrete use-cases will be discussed along with future work directions.
Teacher / speaker
Peter Karsmakers
Dr. Peter Karsmakers received the M.Sc. degree in artificial intelligence in 2004 and the PhD degree from the Department of Electrical Engineering, KU Leuven, in 2010. From 2010 to 2013, he was a post-doctoral researcher in the MOBILAB research team from Thomas More. From 2013 to 2018, he worked as a post-doctoral researcher at KU Leuven where he co-founded the ADVISE research team. Currently, he is an Associate Professor within the Computer Science Department in the DTAI section at KU Leuven and is a member of the Leuven.AI institute. Since 2022, he is a PI of Flanders Make@KU Leuven. His research interests include designing machine learning algorithms that consider application-specific constraints. These can, for example, relate to the computing platform on which the machine learning algorithm will be deployed on, to the need of physical consistency between model variables, to the lack of data annotation or to other application related specifications. He worked on diverse projects, mostly in collaboration with industrial partners, that involve monitoring of both humans as machines using sensors such as microphones, accelerometers and radars.
AI for Time Series
Several research groups in the Flanders AI Research Program conduct world-class research on time series, both in the development of algorithms and tools, as in a wide area of application fields. In a recent poll in the Flanders AI community, ‘time series’ came up as the most wanted topic for future workshops or courses. With this seminar series, we bring together researchers that are interested in, or are conducting research related to, time series. We offer a varied program of national and international speakers.
Klik on the separate webinars below to (re)watch the recordings.
