no code implementations • 28 Feb 2019 • Aleksandr Shevchenko, Anton Osokin
In this paper, we hypothesize that one reason for joint training of deep energy-based models to fail is the incorrect relative normalization of different components in the energy function.