Hardware Implementation of Programmable Neural Networks
A. Klepaczko, A. Napieralski, R. Kielbik, J. M. Moreno
Department of Microelectronics and Computer Science,Technical University of Lodz, PL
Keywords: Neural networks, FPGA, FPSLIC, CHEMFET, SEWING, Water monitoring project, smart sensor
This paper focuses on problems of hardware implementation of neural networks (NN) in the re–programmable structures. New class of these devices, which integrate in one silicon wafer entire microprocessor systems (Systems–on–Chip), facilitates NN construction and their application. The cooperation of Micro–Controller Unit (MCU) and Field Programmable Gate Array (FPGA) helps to overcome space– and interconnection–limitations. Thus, main thesis of this paper which is aimed to be proved states that large multi–layer neural networks are achievable by associating programmable logic array with a micro–controller, what supports space– and speed–efficient designs in comparison to systems realised only in a FPGA device or simulated only by MCU.
Much attention has been devoted to the practical application of the NN in the System for European Water Monitoring (SEWING). Pollution sensors in this system are based on CHEMFET transistors. Each sensor is sensitive on particular ion concentration but other ions can disturb its responses. A possibility of applying NN structure in SEWING for correct interpretation of the chemical sensors responses is the second thesis of this paper.
Main part of the paper describes Multi-Layer Perceptron (MLP) implemented in Atmel’s AT94K FPSLIC. Evaluation board including this chip and connected to PC constitutes an environment able to emulate the system correcting responses of chemical sensors of SEWING project (Figure 1). Presented NN is composed of two layers. Number of neurons in each layer and number of inputs of first layer is programmable. In FPGA included in FPSLIC one neuron is implemented and used repeatedly for each layer as many times as many neurons in this layer is required (Figure 2). This process is controlled by MCU. NN training (weights generation) is performed by means of Matlab’s Neural Network Toolbox on PC.
Tests carried out and described in the paper prove that designed NN fulfils SEWING project requirements. Disturbed responses of CHEMFET-based sensors can be efficiently corrected (Figure 3) in relatively short period of time and using reasonable number of resources.
NSTI Nanotech 2003 Conference Technical Program Abstract