NOSC - Neural Optoelectronic Switch Controller

Project Objectives

(i) Build an optoelectronic based controller, using VCSELs, diffractive optical elements and electronic neurons for a telecommunications switching network which exploits the high parallelism offered by optics.

(ii) Evaluate the performance of the controller circuit with simulated data traffic and use the results to identify the critical operating parameters of the network.

(iii) Extend the proposed controller circuit to include highly compact "smart-pixel" based subsystems which will allow integration and packaging issues to be addressed.

(iv) Extend this optoelectronic approach to other switch topologies and to communications network control.

Project Status: Grant ended January 2001. Work continues.

A Programmable Optoelectronic Neural Network Packet Switch Scheduler

Abstract: A programmable optoelectronic neural network architecture is presented that has been optimized to make routing decisions in both crossbar and banyan packet switch fabrics. Simulation has indicated excellent scalability in this particular application with only a minimal increase in decision time even when problem set size grows by an order of magnitude. Experimental results are presented that demonstrate a high tolerance to both noise and component inconsistencies. An assessment of system performance is made using the common metric of connections per second (CPS).

K. J. Symington, Y. Randle, A. J. Waddie, M. R. Taghizadeh and J. F. Snowdon, "A Programmable Optoelectronic Neural Network Packet Switch Scheduler", OSA Technical Digest of Optics in Computing 2003, Washington, USA, June 2003.

Download the postdeadline paper in PDF format (232kB) or the postdeadline talk in PDF format (2.8MB).

Experimental Implementation of an Optoelectronic Neural Network Scheduler

Abstract: To follow.

A. J. Waddie, Y. Randle, K. J. Symington, J. F. Snowdon and M. R. Taghizadeh, "Experimental Implementation of an Optoelectronic Neural Network Scheduler", accepted for publication in Special Issue on Optical Interconnects, IEEE Journal of Selected Topics in Quantum Electronics.

Download the journal paper in PDF format (To follow).

Application of a Neural Network Demonstrator to Optimise the Positioning of Kings, Knights or Queens in a 8×8 Grid

Abstract: We reveal how one can use an optical neural network to optimise the positioning of Kings, Queens or Knights in such a way that no one are able to take any one in a single move. This project also demonstrates how easy it is to change the interconnects in an optical neural network and prepare it for new tasks. It was expected that this kind of task would not be a problem to solve for this kind of neural network and we demonstrated that this is indeed true. The performance of the optical neural network depends on a set of neural network parameters, labelled A, b and ß, which depend on the neural network itself and the DOE used to define it's interconnects. We found that the ratio A/b is a constant, whose value depends on the DOE, and most values of A and b satisfying that, produce optimal results. It was also established that neuron network performance is independent of ß.

K. Karstad, Application of a Neural Network Demonstrator to Optimise the Positioning of Kings, Knights or Queens in a 8×8 Grid, School of Engineering and Physical Sciences, Heriot-Watt University, Edinburgh, UK, March 2003.

Download the BSc final year report in PDF format (1.26MB).

A Neural-Network Packet Switch Controller: Scalability, Performance, and Network Optimization

Abstract: We examine a novel combination of architecture and algorithm for a packet switch controller that incorporates an experimentally implemented optically interconnected neural network. The network performs scheduling decisions based on incoming packet requests and priorities. We show how and why, by means of simulation, the move from a continuous to a discrete algorithm has improved both network performance and scalability. The system's limitations are examined and conclusions drawn as to its maximum scalability and throughput based on today's technologies.

K. J. Symington, A. J. Waddie, M. R. Taghizadeh and J. F Snowdon, "A Neural-Network Packet Switch Controller: Scalability, Performance, and Network Optimization", IEEE Trans. On Neural Networks, vol. 14, no. 1, pp. 28-34, January 2003.

Download the journal paper in PDF format (922kB)

Demonstrating a Bright Future [Optoelectronic Demonstrators]

Abstract: Micro-optical optoelectronic system demonstrators show that optics in computing interconnection technologies are nearing real-world readiness.

M. R. Taghizadeh and A. J. Waddie, "Demonstrating a Bright Future [Optoelectronic Demonstrators]", IEEE Circuits and Devices Magazine, vol. 18, no. 6, pp. 17-22, November 2002.

Download the article in PDF format (712kB)

Optoelectronic and Optical Device Characteristics for VCSEL-Based Optoelectronic Neural Networks

Abstract: Two dimensional optoelectronic device arrays, especially compact, high efficiency Vertical Cavity Surface Emitting Lasers (VCSELs), have been increasing in array size while decreasing in overall power requirements. This increase in optoelectronic channel density has driven the development of optics-in-computing demonstrators to greater array sizes, such that some of these demonstrator systems have practically realistic processing power.  In this paper we describe the design, construction and successful operation of a 64 neuron optoelectronic Hopfield-type neural network, based around an 8×8 VCSEL array and a diffractive free space optical interconnection. This network is designed to produce near-optimal solutions to a variety of optimisation problems associated with different types of telecommunications switch.

A. J. Waddie, Y. T. Randle and M. R. Taghizadeh, "Optoelectronic and Optical Device Characteristics for VCSEL-Based Optoelectronic Neural Networks", 15th Annual Meeting of the IEEE-LEOS, vol. 2, pp. 720-721, November 2002.

Download the conference proceedings in PDF format (221kB)

Optoelectronic Neural Network Demonstrator: Program and Results

Abstract: The work was divided into three main parts: To write a program in assembly language to run the DSPs in the demonstrator, to write a program in Borland C++ to run the demonstrator from the computer and run the demonstrator and evaluate its performance for scheduling.

Y. Randle, Optoelectronic Neural Network Demonstrator: Program and Results, School of Engineering and Physical Sciences, Heriot-Watt University, Edinburgh, UK, July 2002.

Download the MSc dissertation in PDF format (688kB)

An Optoelectronic Neural Network Packet Switch Scheduler

Abstract: A detailed examination of work to date on the packet switch scheduler including explanation of network convergence. Contains PowerPoint animation.

K. J. Symington, A. J. Waddie, T. Yasue, M. R. Taghizadeh and J. F. Snowdon, "An Optoelectronic Neural Network Packet Switch Scheduler", Invited Talk, Physics Dept., Heriot-Watt University, Edinburgh, UK, January 2001.

Download the oral presentation in PPT format (11.5MB)

High Performance Optoelectronic Neural Network Scheduler

Abstract: A novel optoelectronic architecture is presented that uses a neural network to provide routing decisions for a packet switch. Recent experiments, combined with simulation studies, show that a digital neuron response improves speed and scalability. The next generation switch (scalable to 62 way) is capable of 2.5 million switch configurations per second. The system includes prioritisation and may be configured for a range of applications.

K. J. Symington, A. J. Waddie, T. Yasue, M. R. Taghizadeh and J. F. Snowdon, "High Performance Optoelectronic Neural Network Scheduler", OSA Technical Digest of Optics in Computing 2001, Lake Tahoe, NV, USA, pp. 12-14, January 2001.

Download the conference proceedings in PDF format (155kB) or the oral presentation in PDF format (6.4MB)

Optoelectronic Architecture for a Dual-Functional Neural Switching Optimiser

Abstract: This thesis presents the optical, optoelectronic, and optomechanical design and describes a novel optoelectronic implementation of dual-functional neural networks for switching optimisation. The demonstrator consists of electronic and optical subsystems. The electronic subsystem employs digital signal processors and analogue/digital converters for Hopfield neural networks to optimise decisions on the throughput of crossbar and self-routing switches. The optical subsystem which holds an 8×8 array of vertical-cavity surface-emitting lasers, photodiode detectors, and a diffractive optical element has been mounted on an optomechanically-designed baseplate whose packaging scheme facilitates focus adjustments and mechanical and thermal stability. The free-space optical architecture exhibits a scalability of the neural networks. It can achieve two different functions with minimal realignments of diffractive fan-out elements. The optomechanical package accommodates all the optical components within a miniature space of approximately 12cm×15cm×25cm. Interconnection theory and optoelectronic components are also covered.

T. Yasue, Optoelectronic Architecture for a Dual-Functional Neural Switching Optimiser, Dept. of Physics, Heriot-Watt University, Edinburgh, UK, September 2000.

Download the MPhil dissertation in PDF format (1.2MB).

Optoelectronics and a Neural Network Packet Switch Controller

Abstract: This presentation examines an optoelectronic neural network that can be used to solve a variety of problems. The design and motivation for the system will be discussed and results presented from the first generation system. Performance and optical system scalability issues for current hardware are examined as well as thoughts on design and engineering issues.

K. J. Symington and J. F. Snowdon, "Optoelectronics and a Neural Network Packet Switch Controller", Invited Talk, Department of Electronic and Electrical Engineering, Strathclyde University, Glasgow, UK, August 2000.

Download the oral presentation in PDF format (10.6MB) or the presentation handouts in PDF format (10.7MB)

An Optoelectronic Neural Network Scheduler: Implementation and Operation

Abstract: The optoelectronic scheduler for a packet switch described previously successfully demonstrated the validity of the Hopfield approach to this problem. However, the system itself was physically large as well as requiring constant observation to ensure that the diverse electronic and optical components remained operational and in alignment. The next-generation demonstrator has been designed to facilitate "hands-off" operation of the demonstrator by using an optomechanical baseplate in conjunction with both off-the-shelf and custom designed optical, electronic and optoelectronic subsystems. In this paper we shall report on the progress of each of these elements of the system as well as investigating the limitations on the scalability of the system imposed by the different subsystems.

A. J. Waddie, T. Yasue, K. J. Symington, J. F. Snowdon and M. R. Taghizadeh, "An Optoelectronic Neural Network Scheduler: Implementation and Operation", OSA Technical Digest of Optics in Computing 2000, Quebec City, Canada, pp. 304-310, June 2000.

Download the conference proceedings in PDF format (3.4MB) or the oral presentation in PDF format (11.6MB)

Optoelectronic Neural Networks

Abstract: A general purpose neural network demonstrator is presented along with its application specific predecessor which employs a winner take all strategy to optimise decisions on the throughput of both a crossbar and a banyan packet switching fabric. The problems of high interconnection density in neural networks are solved by using free space optical interconnects which exploit diffractive optical techniques to generate the required interconnection patterns. The design, construction and operation of the general purpose network is discussed along with the fully operational experimental application as a packet switch scheduler which could significantly outperform current state of the art schedulers.

K. J. Symington, J. F. Snowdon, A. J. Waddie, T. Yasue, and M. R. Taghizadeh, "Optoelectronic Neural Networks", IEE Proc. of PREP 2000, Nottingham, UK, pp. 182-187, April 2000.

Download the conference proceedings in PDF format (408kB), the talk in PDF format (5.4MB) or a summary of selected talks in PDF format (223kB)

Neural Network Simulator

Abstract: Programs that simulate the neural networks described by documents on this page. There are various versions of the software, each adding additional functionality. Although the program manual has not been updated since V2.00, the manual provides a good overview of what the program does.

K. J. Symington, Neural Network Simulator, Heriot-Watt University, Edinburgh, UK.

Download the simulator V3.00 in EXE format (446kB), V2.00 in EXE format (304kB), V1.00 in EXE format (267kB) or the documentation for V2.00 in PDF format (3.65MB).

An Optoelectronic Neural Network Scheduler for Packet Switches

Abstract: A novel type of packet scheduler is presented which could significantly out perform current state of the art schedulers. The operation and design of such a scheduler are discussed and a fully operational experimental implementation is described. The scheduler employs a neural network in a winner take all strategy to optimise decisions on the throughput of both a crossbar and a banyan switching fabric. The problems of high interconnection density are solved by the use of a free space optical interconnect which exploits diffractive optical techniques to generate the required interconnection patterns and weights.

R. P. Webb, A. J. Waddie, K. J. Symington, M. R. Taghizadeh and J. F. Snowdon, "An Optoelectronic Neural Network Scheduler for Packet Switches", Applied Optics, vol. 39, no. 5, pp. 788-795, February 2000.

Download the journal paper in PDF format (533kB).

Optoelectronic Neural Networks for Packet Switching

Abstract: At the core of some of the latest generation of internet routers is a hardware switch that transports packets between the line cards. A central scheduler is required to select a set of packets from queues on the line cards that can be connected to the correct outputs simultaneously without blocking. The larger the set chosen, the greater the throughput, but the decision must be made within the cycle time of the switch. This assignment of outputs to inputs subject to constraints imposed by the switch fabric is an example of a resource allocation problem which can be solved by a neural network.

R. P. Webb, A. J. Waddie, K. J. Symington, M. R. Taghizadeh and J. F. Snowdon, "Optoelectronic Neural Networks for Packet Switching", IoP Technical Digest of Quantum Electronics and Photonics 14, Manchester, UK, September 1999.

Download the conference proceedings in PDF format (33kB) or the poster presentation in PDF format (7.8MB).

Solving the Assignment Problem Using Neural Networks

Abstract: Current software systems suffer from an exponential increase in computational complexity when solving a quadratic assignment problem. This document considers the problem and proceeds to propose a solution using the inherent parallelism of a neural network to reduce computation times. A specific example is given, in this case a crossbar switch, onto which problem mapping is demonstrated and a solution given.

K. J. Symington and J. F. Snowdon, Solving the Assignment Problem Using Neural Networks, Department of Physics, Heriot-Watt University, Edinburgh, UK, June 1999.

Download the summary document in PDF format (592kB).

Optoelectronic Neural Networks for Switching

Abstract: Details the optoelectronic packet switch scheduler in graphical format and its associated technologies. First presented at IEEE-LEOS (Scotland) meeting April 1999 where it won best poster prize. Poster by K. J. Symington.

M. R. Taghizadeh, J. F. Snowdon, A. J. Waddie and K. J. Symington, "Optoelectronic Neural Networks for Switching", Topical Meeting of IEEE-LEOS, Heriot-Watt University, Edinburgh, UK, April 1999.

Download the poster presentation in PDF format (7.8MB).

A Neural Network Scheduler for Packet Switches

Abstract: At the core of some of the latest generation of internet routers is a hardware switch that transports packets between the line cards.  A central scheduler is required to select a set of packets from queues on the line cards that can be connected to the correct outputs simultaneously without blocking.  The larger the set chosen, the greater the throughput, but the decision must be made within the cycle time of the switch. This assignment of outputs to inputs subject to constraints imposed by the switch fabric is an example of a resource allocation problem which can be solved by a Hopfield neural network. This paper examines a optoelectronic hardware implementation which can be used to solve the assignment problem.

R. P. Webb, A. J. Waddie, K. J. Symington, M. R. Taghizadeh and J. F. Snowdon, "A Neural Network Scheduler for Packet Switches", OSA Technical Digest of Optics in Computing 1999, Aspen, CO, USA, pp. 193-195, April 1999.

Download the conference proceedings in PDF format (587kB) or the oral presentation in PDF format (4.0MB).

Implementation of an Optoelectronic Neural Network

Abstract: This proposal examines implementation methodologies for an optoelectronic neural network comparing it to current digital, analogue and hybrid systems. After careful examination of available components, a conclusion is made on the desired characteristics of any demonstrator as well as the final project goal.

Keith J. Symington, Implementation of an Optoelectronic Neural Network, Physics Dept., Heriot-Watt University, Edinburgh, UK, December 1998.

Download the implementation proposal in PDF format (5.4MB).

Optoelectronic Neural Networks for Switching

Abstract: Current software systems suffer from an exponential increase in computational complexity when solving a quadratic assignment problem. Such problems exist in today's telecommunication systems as a network tries to rout calls optimally through its switches to minimise blocking. This project considers the problem and proceeds to propose a solution using the inherent parallelism of a neural network to reduce computation times. In conclusion, a hardware implementation is examined which uses free space optical interconnects to reduce circuit complexity and its performance is closely scrutinised.

K. J. Symington, Optoelectronic Neural Networks for Switching, Dept. of Physics, Heriot-Watt University, Edinburgh, UK, September 1998.

Download the MSc dissertation in PDF format (1.12MB).

Smart Pixel Optoelectronic Neural Networks

Abstract: Control tasks beyond the scope of serial processors can be performed using parallel systems such as neural networks. Current neural systems suffer serious interconnection problems on silicon but by employing spatial optics to interconnect neurons this limitation can be overcome. This report examines both the neural network architectures and smart pixel interconnection technologies that can be used to construct them. In conclusion, it examines a sample system and considers the potential applications of a smart pixel optical neural network.

K. J. Symington, "Smart Pixel Optoelectronic Neural Networks", Dept. of Physics, Heriot-Watt University, Edinburgh, UK, April 1998.

Download the MSc literature search in PDF format (2.8MB)

Last Modified 29/07/03 16:13:49.