Offline training for memristor-based neural networks

Neuromorphic systems based on Hardware Neural Networks (HNN) are expected to be an energy-efficient computing architecture for solving complex tasks. Due to the variability common to all nano-electronic devices, HNN success depends on the development of reliable weight storage or mitigation techniques against weight variation. In this manuscript, we propose a neural network training technique to mitigate the impact of device-to-device variation due to conductance imperfections at weight import in offline-learning. To that aim, we propose to add said variation to the weights during training in order to force the network to learn robust computations against that variation. Then, we experiment using a neural network architecture with quantized weights adapted to the design constrains imposed by memristive devices. Finally, we validate our proposal against real-world road traffic data and the MNIST image data set, showing improvements on the classification metrics.