Influence of Recognition Performance on Recurrent Neural Network Using Phase-Change Memory as Synapses

This paper simulates the influence of recognition performance of using a virtual phase-change memory as weight of a recurrent neural network. In the first experiment, a neural network did not learn well due to NaN error caused by write error in phase-change memory. In the second and later experiments, there was no correlation between the write error rate and the validation loss. However, when the learning results were output, higher write error rate caused the less successful the learning.