Comparison of Consolidation Methods for Predictive Learning of Time Series

In environments where various tasks are sequentially given to deep neural networks (DNNs), training methods are needed that enable DNNs to learn the given tasks continuously. A DNN is typically trained by a single dataset, and continuous learning of subsequent datasets causes the problem of catastrophic forgetting. Previous studies have reported results for consolidation learning methods in recognition tasks and reinforcement learning problems. However, those methods were validated on only a few examples of predictive learning for time series. In this study, we applied elastic weight consolidation (EWC) and pseudo-rehearsal to the predictive learning of time series and compared their learning results. Evaluating the latent space after the consolidation learning revealed that the EWC method acquires properties of the pre-training and subsequent datasets with the same distribution, and the pseudo-rehearsal method distinguishes the properties and acquires them with different distributions.