Abstract:
Spiking Neural Networks (SNNs) are a new promising approach for machine learning because they increase the ratio of biological realism, and thus the ability to capture complex data patterns. SNNs belong to the third generation of Artificial Neural Networks (ANNs). In contrast to the first and second generations of ANNs, SNNs deal with spatial-temporal information effectively. Based on SNNs, a new trend in Machine Learning was introduced, the Liquid State Machine (LSM). LSM is randomly and sparsely recurrent SNNs that was introduced by Wolfgang Maass in 2003 as a solution for firmness in training SNNs and Recurrent Neural Networks (RNNs). One of the advantages of using LSM is the ability to handle data streams from input to generate high-dimensional separable outputs, which makes LSM a suitable approach for dynamical systems pattern recognition. Motivated by the compelling capabilities of LSM, this thesis explores different topics in LSM. First, the work harnesses LSM for Emotion Recognition from EEG signals, where we use LSM as an anytime multi-purpose model for identifying Valence, Arousal and Liking. Second, we utilize LSM for Continuous Authentication from mobile devices and we show the benefits of LSM for such purposes. Third, we explore the possible deployments of LSM for the Feature Extraction task from raw data in comparison with Deep Learning (DL) approach, and more specifically the Deep Belief Networks. Fourth, we introduce an Active Liquid States Selection method to effectively read from LSM and we show how this method can reduce the computational requirement while improving the accuracy. Finally, because of the new trends in building hardware-based LSMs that are energy aware, we introduce an Inattentive Neurons Pruning method to rank and prune the uninformative neurons inside the LSM to reduce power consumption and computational requirements, where the method have shown to be able to provide better accuracies in most of benchmarks, while reducing the number of neurons inside the LSM by up to 50percent.
Description:
Thesis. M.E. American University of Beirut. Department of Electrical and Computer Engineering, 2016. ET:6344
Advisor : Dr. Mariette Awad, Associate Professor, Electrical and Computer Engineering ; Committee Members : Dr. Mohamad Adnan Al-Alaoui, Professor, Electrical and Computer Engineering ; Dr. Nikola Kasabov, Professor, KEDRI, Auckland University of Technology.
Includes bibliographical references (leaves 154-161)