Non-adversarial training of Neural SDEs with signature kernel scores

Neural SDEs are continuous-time generative models for sequential data. State-of-the-art performance for irregular time series generation has been previously obtained by training these models adversarially as GANs.