I finally made a coffee and decided to read Neural Ordinary Differential Equations (Chen et al. 2018); which re-formulates neural networks as differential equations so that they are based on continuous domains that can be trained using any ODE solver.
This results in better memory management, less parameters and better model reconstruction under certain circumstances. What a refreshing point of view, with really interesting possibilities. https://arxiv.org/abs/1806.07366
Server run by the main developers of the project It is not focused on any particular niche interest - everyone is welcome as long as you follow our code of conduct!