A New Approach to a Fast Simulation of Spiking Neural Networks

Rene Mayrhofer

A New Approach to a Fast Simulation of Spiking Neural Networks

Rene Mayrhofer

Research Center Hagenberg

Publikation: Typen von Abschlussarbeiten › Masterarbeit / Diplomarbeit

Abstract

Spikende Neuronale Netzwerke werden aufgrund ihrer verbesserten Flexibilität und erhöhten Anzahl von Freiheitsgraden gerne als ein neues Berechnungs-Paradigma angesehen. Sie stellen den direkten Nachfolger der Künstlichen Neuronalen Netzwerke dar. Obwohl die Eigenschaften dieses neuen Typs Neuronaler Netzwerke derzeit nur in begrenztem Masse bekannt sind, ist er dennoch eindeutig leistungsfähiger als sein Vorgänger; ausser der möglichen Simulation Künstlicher Neuronaler Netzwerke in Echtzeit können neue, zuvor unbekannte Berechnungselemente in der Modellierung verwendet werden. Allerdings erfordern aktuelle Implementierung zur Simulation Spikender Neuronaler Netzwerke bisher den Einsatz kontinuierlicher Simulationstechniken, durch die Skalierbarkeit auf grosse Netzwerke mit vielen Neuronen erschweren. Diese Diplomarbeit führt ein neues Modell für Spikende Neuronale Netzwerke ein, welches die Anwendung von schneller, diskreter ereignisbasierter Simulation erlaubt; dadurch entstehen moglicherweise enorme Vorteile in Flexibilität und Skalierbarkeit, ohne die qualitative Berechnungsleistung zu mindern. Das neue Modell wurde ausserdem in einem Plattform-unabhängigen, in Java geschriebenen Prototyp-Simulationsframework implementiert. Durch die ausschliessliche Verwendung diskreter ereignisbasierter Simulation beweist das Framework die Funktionsfähigkeit des neuen Konzepts. Es wurde bereits erfolgreich zur Emulation von Standardtypen Künstlicher Neuronaler Netzwerke sowie zur Simulation eines biologisch inspirierten Filter-Modells eingesetzt. Die Resultate dieser Simulationen werden in folgenden Kapiteln präsentiert und mögliche Richtungen für zukünftige Weiterentwicklungen angegeben. Zusätzlich werden einige erweiterte Techniken bezülich des Einsatzes diskreter ereignisbasierter Simulation angegeben, um die durch das neue Konzept entstandenen Mögichkeiten nutzen zu können.

Originalsprache	Englisch
Publikationsstatus	Veröffentlicht - 2002

Andere Dateien und Links

http://www.mayrhofer.eu.org/downloads/publications/DiplomaThesis-2002-screen.pdf

Zitieren

@mastersthesis{849210c470744b70a79a66e1e7391e5e,

title = "A New Approach to a Fast Simulation of Spiking Neural Networks",

abstract = "Spiking Neural Networks are considered as a new computation paradigm, representing the next generation of Artiﬁcial Neural Networks by offering more ﬂexibility and degrees of freedom for modeling computational elements. Although this type of Neural Networks is rather new and there exists only a vague knowledge about its features, it is clearly more powerful than its predecessor, not only being able to simulate Artiﬁcial Neural Networks in real time but also offering new computational elements that were not available previously. Unfortunately, the simulation of Spiking Neural Networks currently involves the use of continuous simulation techniques which do not scale easily to large networks with many neurons. In this diploma thesis, a new model for Spiking Neural Networks is introduced; it allows the use of fast discrete event simulation techniques and possibly offers enormous advantages in terms of simulation ﬂexibility and scalability without restricting the qualitative computational power. As a proof of concept, the new model has been implemented in a prototype simulation framework, written platform-independently in Java. This simulation framework utilizes solely discrete event simulation and has been successfully used to emulate typical Artiﬁcial Neural Networks and to simulate a biologically inspired ﬁlter model. The results of the conducted example simulations are presented and possible directions for future research are given. Additionally, a few advanced techniques regarding the use of discrete event simulation, which offers some new opportunities, are shortly discussed.",

author = "Rene Mayrhofer",

year = "2002",

language = "English",

}

TY - THES

T1 - A New Approach to a Fast Simulation of Spiking Neural Networks

AU - Mayrhofer, Rene

PY - 2002

Y1 - 2002

N2 - Spiking Neural Networks are considered as a new computation paradigm, representing the next generation of Artiﬁcial Neural Networks by offering more ﬂexibility and degrees of freedom for modeling computational elements. Although this type of Neural Networks is rather new and there exists only a vague knowledge about its features, it is clearly more powerful than its predecessor, not only being able to simulate Artiﬁcial Neural Networks in real time but also offering new computational elements that were not available previously. Unfortunately, the simulation of Spiking Neural Networks currently involves the use of continuous simulation techniques which do not scale easily to large networks with many neurons. In this diploma thesis, a new model for Spiking Neural Networks is introduced; it allows the use of fast discrete event simulation techniques and possibly offers enormous advantages in terms of simulation ﬂexibility and scalability without restricting the qualitative computational power. As a proof of concept, the new model has been implemented in a prototype simulation framework, written platform-independently in Java. This simulation framework utilizes solely discrete event simulation and has been successfully used to emulate typical Artiﬁcial Neural Networks and to simulate a biologically inspired ﬁlter model. The results of the conducted example simulations are presented and possible directions for future research are given. Additionally, a few advanced techniques regarding the use of discrete event simulation, which offers some new opportunities, are shortly discussed.

AB - Spiking Neural Networks are considered as a new computation paradigm, representing the next generation of Artiﬁcial Neural Networks by offering more ﬂexibility and degrees of freedom for modeling computational elements. Although this type of Neural Networks is rather new and there exists only a vague knowledge about its features, it is clearly more powerful than its predecessor, not only being able to simulate Artiﬁcial Neural Networks in real time but also offering new computational elements that were not available previously. Unfortunately, the simulation of Spiking Neural Networks currently involves the use of continuous simulation techniques which do not scale easily to large networks with many neurons. In this diploma thesis, a new model for Spiking Neural Networks is introduced; it allows the use of fast discrete event simulation techniques and possibly offers enormous advantages in terms of simulation ﬂexibility and scalability without restricting the qualitative computational power. As a proof of concept, the new model has been implemented in a prototype simulation framework, written platform-independently in Java. This simulation framework utilizes solely discrete event simulation and has been successfully used to emulate typical Artiﬁcial Neural Networks and to simulate a biologically inspired ﬁlter model. The results of the conducted example simulations are presented and possible directions for future research are given. Additionally, a few advanced techniques regarding the use of discrete event simulation, which offers some new opportunities, are shortly discussed.

UR - http://www.mayrhofer.eu.org/downloads/publications/DiplomaThesis-2002-screen.pdf

M3 - Master's Thesis / Diploma Thesis

ER -

A New Approach to a Fast Simulation of Spiking Neural Networks

Abstract

Andere Dateien und Links

Fingerprint

Zitieren