Pipelined Backpropagation at Scale: Training Large Models without Batches
This work addresses hardware efficiency and scalability for training large models, offering an incremental improvement in asynchronous pipeline parallel algorithms.
The paper tackles the challenge of training large neural networks without large batches by proposing fine-grained Pipelined Backpropagation with methods like Spike Compensation and Linear Weight Prediction to mitigate asynchronicity issues, achieving accuracy matching SGD on CIFAR-10 and ImageNet with batch size one.
New hardware can substantially increase the speed and efficiency of deep neural network training. To guide the development of future hardware architectures, it is pertinent to explore the hardware and machine learning properties of alternative training algorithms. In this work we evaluate the use of small batch, fine-grained Pipelined Backpropagation, an asynchronous pipeline parallel training algorithm that has significant hardware advantages. We introduce two methods, Spike Compensation and Linear Weight Prediction, that effectively mitigate the downsides caused by the asynchronicity of Pipelined Backpropagation and outperform existing techniques in our setting. We show that appropriate normalization and small batch sizes can also aid training. With our methods, fine-grained Pipelined Backpropagation using a batch size of one can match the accuracy of SGD for multiple networks trained on CIFAR-10 and ImageNet. Simple scaling rules allow the use of existing hyperparameters for traditional training without additional tuning.