Neuron Deactivation Scheme for Defect-Tolerant Memristor Neural Networks

As amounts of data generated from countless and ubiquitous IoT sensors are increased very sharply, memristor crossbars can be considered very suitable to edge intelligence hardware due to high energy efficiency of computing, dense and 3D integration, non-volatility, multi-state memory, CMOS compatible fabrication etc. But, due to the limits of immature fabrication technology, the memristor crossbars can have defects such as stuck-at-faults. To compensate for malfunction of neural networks caused from the fabrication-related defects, in this paper, a simple neuron deactivation scheme is reviewed and analyzed for maximizing its capability to compensate for the neural network’s performance degradation due to the memristor defects. The column deactivation scheme can be particularly useful for the edge intelligence hardware, because the defect map occupying a large amount of memory is not needed during the training. Moreover, the direct mapping from the calculated synaptic weights to the memristor crossbar can save the retraining time required for the defect-aware training scheme.