Special ACM Seminar

Neural stochastic differential equations (Neural SDEs) have proven to be powerful continuous-time generative models that leverage neural networks to parameterize the drift and diffusion functions of stochastic differential equations (SDEs). These models have achieved state-of-the-art performance in generating multivariate time series through adversarial training as GANs. In this talk, I will introduce Neural stochastic partial differential equations (Neural SPDEs), which extend the capabilities of Neural SDEs to model spatio-temporal dynamics in continuous space-time. I will start by explaining how the Neural SPDE model parametrizes the mild solution of an SPDE and then discuss how it can be trained by minimizing suitable scoring rules on path space based on signature kernels.