This study represents the fault tolerance analysis and recovery of multilayer perceptron model used for the pattern recognition. In this paper, we have investigated the faults that may occur in hardware implementations of the model. The effect of two kinds of faults on network performance is simulated and analyzed by modeling the faults and the neural network in software. We will consider two different types of faults-1. stuck-at-faults, 2. faults due to damaged connections between neurons. In case of stuck-at-faults, we considered three kinds of fault-1. stuck-at-0, 2. stuck-at-0.5, and 3. stuck-at-1. In case of faults due to damaged connections between neurons, we considered two kinds of faults-1. reduced connection weights, 2. zero connection weights. We have investigated the output layer neurons' output affected by faults. We found that the output is related with the connection weights with positive sign and those with negative sign. And we found that the damaged neuron can be recovered by magnifying both connection weights with positive sign and those with negative sign.
The copyright of the original papers published on this site belongs to IEICE. Unauthorized use of the original or translated papers is prohibited. See IEICE Provisions on Copyright for details.
Copy
In-Jung PARK, Yong-Joo CHUNG, "Fault Analysis and Recovery of Multilayer Perceptron Model" in IEICE TRANSACTIONS on Fundamentals,
vol. E74-A, no. 10, pp. 3092-3097, October 1991, doi: .
Abstract: This study represents the fault tolerance analysis and recovery of multilayer perceptron model used for the pattern recognition. In this paper, we have investigated the faults that may occur in hardware implementations of the model. The effect of two kinds of faults on network performance is simulated and analyzed by modeling the faults and the neural network in software. We will consider two different types of faults-1. stuck-at-faults, 2. faults due to damaged connections between neurons. In case of stuck-at-faults, we considered three kinds of fault-1. stuck-at-0, 2. stuck-at-0.5, and 3. stuck-at-1. In case of faults due to damaged connections between neurons, we considered two kinds of faults-1. reduced connection weights, 2. zero connection weights. We have investigated the output layer neurons' output affected by faults. We found that the output is related with the connection weights with positive sign and those with negative sign. And we found that the damaged neuron can be recovered by magnifying both connection weights with positive sign and those with negative sign.
URL: https://global.ieice.org/en_transactions/fundamentals/10.1587/e74-a_10_3092/_p
Copy
@ARTICLE{e74-a_10_3092,
author={In-Jung PARK, Yong-Joo CHUNG, },
journal={IEICE TRANSACTIONS on Fundamentals},
title={Fault Analysis and Recovery of Multilayer Perceptron Model},
year={1991},
volume={E74-A},
number={10},
pages={3092-3097},
abstract={This study represents the fault tolerance analysis and recovery of multilayer perceptron model used for the pattern recognition. In this paper, we have investigated the faults that may occur in hardware implementations of the model. The effect of two kinds of faults on network performance is simulated and analyzed by modeling the faults and the neural network in software. We will consider two different types of faults-1. stuck-at-faults, 2. faults due to damaged connections between neurons. In case of stuck-at-faults, we considered three kinds of fault-1. stuck-at-0, 2. stuck-at-0.5, and 3. stuck-at-1. In case of faults due to damaged connections between neurons, we considered two kinds of faults-1. reduced connection weights, 2. zero connection weights. We have investigated the output layer neurons' output affected by faults. We found that the output is related with the connection weights with positive sign and those with negative sign. And we found that the damaged neuron can be recovered by magnifying both connection weights with positive sign and those with negative sign.},
keywords={},
doi={},
ISSN={},
month={October},}
Copy
TY - JOUR
TI - Fault Analysis and Recovery of Multilayer Perceptron Model
T2 - IEICE TRANSACTIONS on Fundamentals
SP - 3092
EP - 3097
AU - In-Jung PARK
AU - Yong-Joo CHUNG
PY - 1991
DO -
JO - IEICE TRANSACTIONS on Fundamentals
SN -
VL - E74-A
IS - 10
JA - IEICE TRANSACTIONS on Fundamentals
Y1 - October 1991
AB - This study represents the fault tolerance analysis and recovery of multilayer perceptron model used for the pattern recognition. In this paper, we have investigated the faults that may occur in hardware implementations of the model. The effect of two kinds of faults on network performance is simulated and analyzed by modeling the faults and the neural network in software. We will consider two different types of faults-1. stuck-at-faults, 2. faults due to damaged connections between neurons. In case of stuck-at-faults, we considered three kinds of fault-1. stuck-at-0, 2. stuck-at-0.5, and 3. stuck-at-1. In case of faults due to damaged connections between neurons, we considered two kinds of faults-1. reduced connection weights, 2. zero connection weights. We have investigated the output layer neurons' output affected by faults. We found that the output is related with the connection weights with positive sign and those with negative sign. And we found that the damaged neuron can be recovered by magnifying both connection weights with positive sign and those with negative sign.
ER -