Magnetoresistive RAM for error resilient XNOR-Nets
This work addresses power consumption and hardware compatibility issues for machine learning systems, particularly in enabling more efficient memory technologies like MRAM, though it is incremental in nature.
The study investigated the resilience of Binarized Convolutional Neural Networks to activation errors during training, finding that bit error rates of a few percent did not degrade test accuracy in most cases, except for AlexNet on ImageNet, and that tolerance increased with smaller subsets.
We trained three Binarized Convolutional Neural Network architectures (LeNet-4, Network-In-Network, AlexNet) on a variety of datasets (MNIST, CIFAR-10, CIFAR-100, extended SVHN, ImageNet) using error-prone activations and tested them without errors to study the resilience of the training process. With the exception of the AlexNet when trained on the ImageNet dataset, we found that Bit Error Rates of a few percent during training do not degrade the test accuracy. Furthermore, by training the AlexNet on progressively smaller subsets of ImageNet classes, we observed increasing tolerance to activation errors. The ability to operate with high BERs is critical for reducing power consumption in existing hardware and for facilitating emerging memory technologies. We discuss how operating at moderate BER can enable Magnetoresistive RAM with higher endurance, speed and density.