Fault-tolerant artificial neural networks
J.H. Kim, Chidchanok Lursinsap, S. Park · 2002
Summary form only given, as follows. Self-recovery methods in artificial neural networks (ANNs) implemented on a digital VLSI chip were investigated. Fault tolerance is the potential benefit of ANNs that extends beyond the high computation rates facilitated by the massive parallelism. If a faulty neuron or a faulty link occurs in ANNs implemented on a VLSI chip, typically ANNs no longer classify all inputs correctly. The ability of ANNs to achieve fault tolerance is not inherent, but must be built in. Also, the built-in fault-tolerant mechanism must be practical and efficient enough for a VLSI chip implementation. A partial relearning scheme was proposed to achieve fault tolerance. The scheme was applied to only a single neuron level, not entire networks. Therefore, the execution speed of the partial relearning will be much faster than that of the normal learning. Furthermore, the partial relearning can be executed in a parallel fashion.>