Abstract:

We consider a typical problem of machine learning - the reconstruction
of probability distributions of observed data. We introduce the
so-called gradient conjugate prior (GCP) update and study the induced
dynamical system. We will explain the dynamics of the parameters and
show how one can use insights from the dynamical behavior to recover the
ground truth distribution in a way that is more robust against outliers.
The developed approach also carries over to neural networks.
This is joint work with Pavel Gurevich.