Supplementary Material Causes and Effects of Unanticipated Numerical Deviations in Neural Network Inference Frameworks

For CIFAR-10, model Cifar10-small reaches 53.18 % accuracy, and the Cifar10-R18 reaches 60.25 % accuracy. These accuracies are not competitive with the state of the art, but sufficiently better than random guessing. We can safely assume that the kernels learn meaningful weights. Experiment samples We process three samples for each of our models to measure the consistency of our results. The first sample is the first test sample (for simplicity); we additionally use a sample from a different class (sample index 1 for CIFAR-10, and index 6 for Deep Weeds), a sample from the same class as the first sample is also used (index 6 for CIFAR-10, and index 1 for Deep Weeds). All sample indexes refer to the unshuffled test set of the respective dataset.