Reservoir Computing for Learning in Structured Domains

The study of learning models for direct processing complex data structures has gained an increasing interest within the Machine Learning (ML) community during the last decades. In this concern, efficiency, effectiveness and adaptivity of the ML models on large classes of data structures represent challenging and open research issues. The paradigm under consideration is Reservoir Computing (RC), a novel and extremely efficient methodology for modeling Recurrent Neural Networks (RNN) for adaptive sequence processing. RC comprises a number of different neural models, among which the Echo State Network (ESN) probably represents the most popular, used and studied one. Another research area of interest is represented by Recursive Neural Networks (RecNNs), constituting a class of neural network models recently proposed for dealing with hierarchical data structures directly. In this thesis the RC paradigm is investigated and suitably generalized in order to approach the problems arising from learning in structured domains. The research studies described in this thesis cover classes of data structures characterized by increasing complexity, from sequences, to trees and graphs structures. Accordingly, the research focus goes progressively from the analysis of standard ESNs for sequence processing, to the development of new models for trees and graphs structured domains. The analysis of ESNs for sequence processing addresses the interesting problem of identifying and characterizing the relevant factors which influence the reservoir dynamics and the ESN performance. Promising applications of ESNs in the emerging field of Ambient Assisted Living are also presented and discussed. Moving towards highly structured data representations, the ESN model is extended to deal with complex structures directly, resulting in the proposed TreeESN, which is suitable for domains comprising hierarchical structures, and Graph-ESN, which generalizes the approach to a large class of cyclic/acyclic directed/undirected labeled graphs. TreeESNs and GraphESNs represent both novel RC models for structured data and extremely efficient approaches for modeling RecNNs, eventually contributing to the definition of an RC framework for learning in structured domains. The problem of adaptively exploiting the state space in GraphESNs is also investigated, with specific regard to tasks in which input graphs are required to be mapped into flat vectorial outputs, resulting in the GraphESN-wnn and GraphESN-NG models. As a further point, the generalization performance of the proposed models is evaluated considering both artificial and complex real-world tasks from different application domains, including Chemistry, Toxicology and Document Processing.