In unsupervised training of deep neural networks one often seeks to
reproduce the probability distribution of the dataset. In this talk
I'll discuss a different goal which is to make the log-probability of
the coarse grained dataset as local as possible. I'll argue that this
goal mappable to a certain information-bottleneck problem and also to
the renormalization-group (RG) technique common in physics.
Furthermore I'll present recent analytical results supporting this
explicit link between the information-bottleneck approach and RG and give an outlook on some potential applications in machine learning and
physics.