Javascript must be enabled for the correct page display

Nets In Space, Spatial Design of a Modular Neural Network by Reconfiguration of an FPGA

Alberts, R. (2002) Nets In Space, Spatial Design of a Modular Neural Network by Reconfiguration of an FPGA. Master's Thesis / Essay, Computing Science.

Infor_Ma_2002_RAlberts.CV.pdf - Published Version

Download (2MB) | Preview


The digital implementation of neural networks has never become really popular. The synapses seem too numerous to be physically shaped and therefore more easily handled in software. Further, their operation requires a multiplication, which is electrically easy but logically cumbersome. The idea of many simple nodes, that in combination produce a complicated function, seemed like a fairy tale. But micro-electronic technology has changed this picture drastically. At the start of the silicon era, the lack of integration drove towards a temporal computing style, whereby many tasks were scheduled for the optimal use ofjust a few resources. But the level of integration rose faster than the design efficiency. This creates a ,,Productivity Gap": in current technology we have more resources available than we can optimally use. It is suggested that in contrast to the past we can now utilize a spatial computing style, whereby few tasks are roaming over many resources. In a typical spatial device like a Field-Programmable Gate-Array (FPGA) we find configurable interconnect & logic, mixed with memory and arithmetic macros. Configuration blocks take the role of program segments and re-configuration schemes replace temporal scheduling. While adequate CAD tools are still lacking, the first challenge is to envision what this new computing style has to offer. This is best learned by experimentation. Hence, this thesis looks into the potential of modern FPGA devices to implement neural networks. A neural network can be constructed from SRAM and multiplier macros, glued together by the Configurable Logic & Interconnect Blocks. As the implementation of a complete network, this has too much similarity with temporal computing; as the implementation of a single neuron we still have the classical size problems. Here, we investigate the modular neural network: many small networks that are dynamically configured into a virtual large one. We show that this concept is scalable, utilizes the resources efficiently and allows for a high-level behavioral abstraction during design. Hereby it illustrates a number of potential advantages of the spatial computing style.

Item Type: Thesis (Master's Thesis / Essay)
Degree programme: Computing Science
Thesis type: Master's Thesis / Essay
Language: English
Date Deposited: 15 Feb 2018 07:29
Last Modified: 15 Feb 2018 07:29

Actions (login required)

View Item View Item