ETD system

Electronic theses and dissertations repository


Tesi etd-12022011-181151

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