Zum Hauptinhalt springen


Generative deep learning, dynamic modeling & explainable AI: meaning for app. biostatistics | Prof. Harald Binder (Uni Freiburg)

Beginn: 13. Juli 2021, 16:00 Uhr
Ende: 13. Juli 2021, 17:30 Uhr
Kategorie: Kolloquien Mathematik , Fachbereich 3 - Kolloquien , Fachbereich 3 - Veranstaltungen

In the last decade, the deep learning community has made considerable
progress in the modeling of complex data, in particular with image data.
Artificial neural networks are a core building block that allows for flexible
non-linear functions in the corresponding models. Furthermore, deep
generative approaches can be used to model the joint distribution of
variables, and to subsequently draw synthetic observations. Recent
techniques also allow to combine neural networks with differential equations
for dynamic modeling. Many of these approaches have a black-box
character, i.e. do not meet the requirement of interpretability needed in
typical application settings of applied biostatisticians. Yet, an “explainable
artificial intelligence (AI)” movement is currently emerging that strives to fill
this gap. I will use exemplary applications to illustrate how corresponding
techniques can be used to perform complex modeling tasks while still
obtaining interpretability. In a first example, I will show how synthetic
observations from deep generative models can be useful for uncovering
structure in single cell gene expression data. In a second example, I will
demonstrate how a combination with differential equations allows to
incorporate domain knowledge, and to characterize individual dynamic
trajectories. Together, these examples will illustrate how applied
biostatisticians can extend their toolbox with developments from the deep
learning community, and also contribute their own perspective.