The balance between feature learning and collapse in dynamical systems-based generative models
This theoretical paper uses dynamical systems and operator theory to analyze gradient flow for a general class of stochastic differential equations used in continuous-time generative models. It links mode collapse to low...