SparseOpt: Addressing Normalization-induced Gradient Skew in Sparse Training
For practitioners of sparse neural network training, this work identifies and mitigates a key bottleneck (normalization-induced gradient skew) that previously hindered DST from being practically competitive with dense training.
Dynamic Sparse Training (DST) methods suffer from slow convergence due to normalization-induced gradient skew. SparseOpt, a sparsity-aware optimizer, addresses this and achieves faster convergence and improved generalization on ResNet models across CIFAR-100 and ImageNet.
Dynamic Sparse Training (DST) methods train neural networks by maintaining sparsity while dynamically adapting the network topology. Despite the promise of reduced computation, DST methods converge significantly slower than dense training, often requiring comparable training time to achieve similar accuracy. We demonstrate both analytically and empirically that Batch Normalization (BN) adversely affects sparse training, and propose SparseOpt, a sparsity-aware optimizer, to address this. Experiments on ResNet models across CIFAR-100 and ImageNet demonstrate consistently faster convergence and improved generalization with our proposed method. Our work highlights the limitations of current normalization layers in sparse training and provides the first systematic study of the interaction between Batch Normalization, sparse layers, and DST, taking a significant step toward making DST practically competitive with dense training.