Low Area Overhead In-Situ Training Approach For Memristor-Based Classifier

We propose combination of "dropout" and "Manhattan Rule" training algorithms for memristive crossbar neural networks to reduce circuit area overhead of in -situ training. Using accurate phenomenological model of memristive devices, we show that such combination allows achieving 0.7% misclassification rate on the MNIST benchmark, which is comparable to the best reported results. At the same time, the considered training approach allows reducing the size of memory circuits, the largest area overhead component, which is required to store intermediate weight adjustments during training, by as much as 40% at 16% longer training time as compared to the baseline crossbar circuit compatible "Manhattan Rule" training. The further reduction of the memory circuit area overhead is possible but at the expense of inferior classification performance.