Reliability of Neural Networks Based on Spintronic Neurons

Spintronic technology is emerging as a direction for the hardware implementation of neurons and synapses neuromorphic architectures. In particular, single spintronic device can be used to implement nonlinear activation function neurons. Here, we present how with sigmoidal rectified linear unit (ReLU)-like functions. We then perform numerical experiment showing reliability neural networks made by neurons, all having different functions emulate device-to-device variations in possible network. Therefore, consider “vanilla'' network implemented recognize categories Mixed National Institute Standards Technology database, show an average accuracy 98.87% test dataset, which very close 98.89% obtained ideal case (all have same sigmoid function). Similar results are ReLU-like function.