It was truly an honour to host two speakers today:
Marnix Van Soom is a PhD student at the cognitive AI research group housed in the Artificial Intelligence Lab of Vrije Universiteit Brussel.
Listen to your Prior: Gaussian Processes for Modeling Speech
Abstract: A reliable and accurate white-box model of speech is still something of a holy grail in phonetics, forensics, and clinical medicine. The difficulty is that fitting speech models to microphone recordings is a blind inverse problem where both source signal and the filter operator must be inferred simultaneously from a single waveform.
To make that possible we must lean hard on our prior information to constrain the set of possible solutions: they must adhere to what we already know to be true.
I will show how Gaussian processes let us bake such prior knowledge directly into the model by learning a surrogate kernel from simulated data, and how this improves reliability and accuracy at inference time.
[slides]
Simon Lejoly is a PhD student in the HuMaLearn team of the NaDi Institute in the Computer Science Faculty of the University of Namur.
Efficient Multi-task Gaussian Processes to Study the Atmosphere of Venus
Abstract: The field of spatial aeronomy explores the atmospheres of other planets, aligning models with observational data. Temperature profiles are a common observational product, providing estimations of temperatures at specific altitudes. Given the characteristic shape of temperature profiles and the inherent uncertainty in atmospheric measurements, Gaussian processes (GPs) offer promising tools in such context. This talk will cover various adaptations of GPs for atmospheric datasets, addressing both modeling considerations and computational efficiency.
[slides]
