Improving Layer-wise Adaptive Rate Methods using Trust Ratio Clipping
This work provides an incremental improvement for researchers and practitioners working on large-batch training of neural networks, aiming to maintain accuracy while reducing training time.
The paper addresses the challenge of maintaining accuracy in large-batch neural network training, which typically reduces training time but degrades performance. The authors propose LAMBC, a variant of LAMB, that uses trust ratio clipping to stabilize magnitudes and prevent extreme values, demonstrating improved performance on ImageNet and CIFAR-10.
Training neural networks with large batch is of fundamental significance to deep learning. Large batch training remarkably reduces the amount of training time but has difficulties in maintaining accuracy. Recent works have put forward optimization methods such as LARS and LAMB to tackle this issue through adaptive layer-wise optimization using trust ratios. Though prevailing, such methods are observed to still suffer from unstable and extreme trust ratios which degrades performance. In this paper, we propose a new variant of LAMB, called LAMBC, which employs trust ratio clipping to stabilize its magnitude and prevent extreme values. We conducted experiments on image classification tasks such as ImageNet and CIFAR-10 and our empirical results demonstrate promising improvements across different batch sizes.