Measurement-driven neural-network training for integrated magnetic tunnel junction arrays

Defective devices can severely impact the performance of hardware-based neural networks, in particular resistive crossbar arrays. This study introduces a network training approach that reduces the influence of defective devices, maintaining inference accuracy. The authors demonstrate this approach on a set of dies each containing a crossbar array consisting of 20,000 magnetic tunnel junction devices. They also develop a generalized approach using the statistics of defects and demonstrate similar performance on all dies. These results translate to a manufacturing setting where millions of dies with possible defects are produced, but the performance of even subpar chips can be guaranteed.