Committee Machines -- Universal Method to Deal with Non-Idealities in RRAM-Based Neural Networks

Artificial neural networks (ANNs) are notoriously power- and time-consuming when implemented on conventional von Neumann computing systems. Recent years have seen an emergence of research in hardware that strives to break the bottleneck of von Neumann architecture; namely to bring memory and computing closer together. One of the most often suggested solutions is the physical implementation of ANNs in which their synaptic weights are realised with analogue resistive devices, such as resistive random-access memory (RRAM). However, various device- and system-level non-idealities usually prevent these physical implementations from achieving high inference accuracy. We suggest applying a well-known concept in computer science -- committee machine (CM) -- in the context of RRAM-based neural networks. Using simulations and experimental data from three different types of RRAM devices, we show that CMs employing ensemble averaging can successfully increase inference accuracy in physically implemented neural networks that suffer from faulty devices, programming non-linearities, random telegraph noise, cycle-to-cycle variability and line resistance. We show that the accuracy can be improved even without increasing the number of devices.