Optimization of Reservoir Adaptation for Multivariate Time Series Classification
With the increasing need for real-time human health monitoring and the advent of activity tracking devices in our daily lives, temporal data is more available than ever. Time series are encountered in many real-world applications ranging from electronic health records (EHR) and human activity recognition to human biosignal classification. One powerful model for time series learning tasks is Reservoir Computing (RC). Due to its recurrent nature, it is capable of encoding temporal dependencies of time series and because of its unique architec- ture (fixed recurrent layer), it is much faster to train compared to state-of-the-art recurrent models. In this dissertation, we propose an Adaptive Reservoir training method for Echo State Networks (ESN). This method enables the reservoir to adapt to the training task only when it results in more discriminative representations for the input time series. In addition as an example of real-world time series applications, we propose an athletic performance moni- toring framework. Finally, we conclude with an in-depth region-based analysis of COVID-19 pandemic events not only because it is another time series related task but since as this dissertation is being written it has affected us all and has taken many lives.