Multimodal Meta-Learning for Time Series Regression

Recent work has shown the efficiency of deep learning models such as Fully Convolutional Networks (FCN) or Recurrent Neural (RNN) to deal with Time Series Regression (TSR) problems. These sometimes need a lot data be able generalize, yet time series are not long enough learn patterns. Therefore, it is important make use information across improve learning. In this paper, we will explore idea using meta-learning for quickly adapting model parameters new short-history by modifying original Model Agnostic Meta-Learning (MAML) []. Moreover, based on prior multimodal MAML [], propose method conditioning through an auxiliary network that encodes global extract meta-features. Finally, apply different domains, pollution measurements, heart-rate sensors, and electrical battery data. We show empirically our proposed learns TSR few fast outperforms baselines in 9 12 experiments.