Avoiding Bias While Pruning Neural Networks: The Case of Image Classification

Deep neural networks (DNNs) often need to be reduced in size for implementation in customized hardware for applications and products. The fact that DNNs are almost always significantly overparameterized means that there is much room to reduce the size of these networks without adversely impacting their overall performance for most tasks. In this paper, we study whether pruning of DNNs introduces bias in the classification task above and beyond the trained but unpruned model and if it does, when and how to eliminate it. We define bias as the relative classification error frequency of each class before and after pruning, as measured by a χ<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> homogeneity test. To illustrate our methodology, we focus on the image classification task using popular datasets such as MNIST, Fashion-MNIST, and CIFAR10 using high-performance DNNs such as VGG11, AlexNet, Transformers, customized CNNs and fully-connected networks and compare the error frequencies before and after pruning using the categorical χ<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> test. We observe that standard pruning schemes, even when they maintain the overall accuracies of the full DNN, can introduce bias. We show that it is possible to reduce such bias to small and negligible values using the framework of a truncated lottery ticket scheme, which we introduce and evaluate.