This page documents the reference implementation used in the following paper on online training of vanilla RNNs using UORO for medical time-series forecasting.

We provide an implementation of UORO using closed-form simplifications for quantities appearing in the loss gradient estimation for vanilla RNNs. We experimentally compare UORO with several baselines, including Real-Time Recurrent Learning (RTRL), in a respiratory motion-trace forecasting problem relevant to radiotherapy. We take into account hyperparameter tuning, inference time, prediction oscillations, and the influence of the prediction horizon h and ground-truth signal irregularity. UORO yielded lower horizon-averaged RMSE, nRMSE, and maximum error than the other algorithms, and was generally the most accurate method for medium-to-high values of h. It also exhibited satisfactory robustness to input-signal unsteadiness, limited step-to-step prediction fluctuations, and relatively low processing time.