Paper:
A Comparison Between 3 Different GSN Model Hardware Implementations with the Appliance of an ANN Fast Prototyping System
SIMÕES, E.V., UEBEL, L.F., BARONE, D.A.C
World Congress on Neural Networks, Washington, D.C., USA, July, 1995, CDROM proceedings.
download compressed postscript or download PDF.
Abstract:
This work presents a comparison between two
strategies for ANN (Artificial Neural Network) hardware design: VLSI-Full Custom approach and FPGA. For that reason,
three alternatives will be described. The first one is the DIANNE integrated circuit.
It is a full digital-asynchronous implementation of the GSN Boolean neural network, proposed by Edson
Filho et al. The DIANNE circuit consists of a fixed architecture, presenting 4 GSN pyramids with
60 neurons. The learning mode is not included in DIANNE because it occupies a large
silicon area. Therefore, this circuit can only operate in the recall mode. After the software
tool has trained a neuron net, the weights are stored in internal RAM memories that configure
the chip to recognize the trained patterns. The relationship between area, performance and
number of neurons presents a good solution that improves system speed ( compared to full software
implementation) more than 300%.
The second alternative presented in this article is a fast prototyping system for
Boolean neural network design within a proper FPGA matrix, named FLECHA. In such context,
considering only the GSN Boolean model, the software system enables the designer to define
the structure of the neural network and the pattern to be learned. It also performs the
simulation of the ANN, helping to choose a particular architecture and deliver a VHDL
description, which is taken as input to a FPGA prototyping tool. This object-oriented
software tool was developed according to neural networks and genetic algorithms
techniques. It programs the FLECHA matrix to work according the trained neural network. The
same VHDL
description of the trained neurons can be implemented by some commercial FPGA
design tools, such as the ALTERA MAXII+PLUS. This is the third alternative presented
in this paper.
This article contains a brief description of how the proposed ANN design system
works, as well as a brief presentation of the DIANNE circuit and the FLECHA matrix, and
finally, the results obtained from the prototyping of some GSN neural network applications
within the DIANNE circuit, the FLECHA matrix and the ALTERA commercial EPLD.
Eduardo do Valle Simões,
simoes@icmc.usp.br
Back to Eduardo do Valle Simoes Web Page