Learnable Path in Neural Controlled Differential Equations

Neural controlled differential equations (NCDEs), which are continuous analogues to recurrent neural networks (RNNs), a specialized model in (irregular) time-series processing. In comparison with similar models, e.g., ordinary (NODEs), the key distinctive characteristics of NCDEs i) adoption path created by an interpolation algorithm from each raw discrete sample and ii) Riemann--Stieltjes integral. It is makes be RNNs. However, use existing algorithms create path, unclear whether they can optimal path. To this end, we present method generate another latent (rather than relying on algorithms), identical learning appropriate method. We design encoder-decoder module based NODEs, special training for it. Our shows best performance both classification forecasting.