Fall 2019 Postdoc Seminar Series
October 9, 2019 –
Talk Title: Representation Learning via Disentangled Variational Autoencoders
Speaker: Matthias Sachs, SAMSI Postdoctoral Fellow and Duke Researcher
In this talk I will present some insight into representation learning in the context of the variational autoencoder framework which I gained during a SAMSI-industry collaboration with a biostatistics group at the Bayer corporation. I will briefly explain the general idea behind a Variational Autoencoder (VAE) motivating the construction of a VAE by the problem of parameterizing a complex generative model and continue with the discussion of VAE approaches aiming at a disentangled code representation (i.e., a code representation whose components correspond to independent factors of the parametrized generative model). I will close the talk by presenting results obtained by applying the discussed techniques to patients' monitor data in the context of the industry collaboration.
No references provided at this time