Hyperparameter specialization in the Hierarchical-Task Reservoir
The Reservoir Computing framework presents a more efficient approach to training Recurrent Neural Networks. A classic implementation of this framework is the Echo State Network. In addition to being easier to train, Echo State Networks can display a variety of dynamics, which are controlled by the network hyperparameters. When working with complex multiscale sequential data the capabilities of singular Echo State Networks can prove insufficient. In those cases, Hierarchical Echo State Networks are the better choice. Hierarchical Echo State Networks display richer dynamics by having different timescales on each layer of the model. Recently, a novel Hierarchical Echo State Network, called the Hierarchical-Task Reservoir, has been introduced. Its key feature is abstraction. The model’s task is divided into sub-tasks. Every layer in the Hierarchical-Task Reservoir performs a different sub-task as opposed to ”vanilla” Hierarchical Echo State Networks, where each layer has the same task. The analysis of the Hierarchical-Task Reservoir’s performance showed that the model has higher accuracy with respect to Hierarchical Echo State Networks with no task abstraction. This thesis focuses on investigating how the Hierarchical-Task Reservoir is able to achieve better performance. The novel model is analyzed from the perspective of its hyperparameters. The goal is to find out how the parameters control the model accuracy and which parameters have the most impact. Additionally, the effect of the abstraction feature on the parameter values is evaluated, as well as the weight of the feature’s contribution to the model accuracy, in comparison to the contribution of hyperparameters. It is concluded that the model performs best with optimal parameter values and that the abstraction feature has most effect on the model dynamics.
Faculteit der Sociale Wetenschappen