Weight value testing in artificial neural networks
Lee A. Belfore · 2002
This paper presents a method for testing ANN interconnection weight values in programmable ANNs. The methodology requires separate weights for excitatory and inhibitory contributions to a neuron's activation. In the test the excitatory and inhibitory weights are set to be equal, producing no net contribution to the activation. Any faults in the weight values result in an imbalance that is amplified by the neuron activation function and is easily detectable at the neuron output. Analysis is performed on the detection approach that includes fault detection thresholds as a function of neuron gain, network size, and weight perturbation. A methodology is outlined that uses the proposed testing approach followed by example test results.