On the numerical reliability of nonsmooth autodiff: a MaxPool case study
It addresses numerical instability issues in AD for practitioners using nonsmooth operations in deep learning, though it is incremental as it focuses on a specific case study.
This paper investigates the numerical reliability of automatic differentiation (AD) for neural networks with the nonsmooth MaxPool operation, finding that Jacobians with lower norms improve stability and test accuracy, while higher norms cause instability and performance drops, with effects mitigated by techniques like batch normalization or higher precision.
This paper considers the reliability of automatic differentiation (AD) for neural networks involving the nonsmooth MaxPool operation. We investigate the behavior of AD across different precision levels (16, 32, 64 bits) and convolutional architectures (LeNet, VGG, and ResNet) on various datasets (MNIST, CIFAR10, SVHN, and ImageNet). Although AD can be incorrect, recent research has shown that it coincides with the derivative almost everywhere, even in the presence of nonsmooth operations (such as MaxPool and ReLU). On the other hand, in practice, AD operates with floating-point numbers (not real numbers), and there is, therefore, a need to explore subsets on which AD can be numerically incorrect. These subsets include a bifurcation zone (where AD is incorrect over reals) and a compensation zone (where AD is incorrect over floating-point numbers but correct over reals). Using SGD for the training process, we study the impact of different choices of the nonsmooth Jacobian for the MaxPool function on the precision of 16 and 32 bits. These findings suggest that nonsmooth MaxPool Jacobians with lower norms help maintain stable and efficient test accuracy, whereas those with higher norms can result in instability and decreased performance. We also observe that the influence of MaxPool's nonsmooth Jacobians on learning can be reduced by using batch normalization, Adam-like optimizers, or increasing the precision level.