signSGD with Majority Vote is Communication Efficient And Fault Tolerant
This work addresses communication efficiency and fault tolerance for distributed deep learning, offering a practical solution with proven convergence and adversarial robustness, though it is incremental in building upon existing sign-based methods.
The paper tackles the communication bottleneck and fault tolerance issues in distributed neural network training by proposing signSGD with majority vote, which reduces communication by 32x per iteration and proves robustness against up to 50% adversarial workers, while achieving a 25% reduction in training time for ResNet50 on ImageNet with 15 machines.
Training neural networks on large datasets can be accelerated by distributing the workload over a network of machines. As datasets grow ever larger, networks of hundreds or thousands of machines become economically viable. The time cost of communicating gradients limits the effectiveness of using such large machine counts, as may the increased chance of network faults. We explore a particularly simple algorithm for robust, communication-efficient learning---signSGD. Workers transmit only the sign of their gradient vector to a server, and the overall update is decided by a majority vote. This algorithm uses $32\times$ less communication per iteration than full-precision, distributed SGD. Under natural conditions verified by experiment, we prove that signSGD converges in the large and mini-batch settings, establishing convergence for a parameter regime of Adam as a byproduct. Aggregating sign gradients by majority vote means that no individual worker has too much power. We prove that unlike SGD, majority vote is robust when up to 50% of workers behave adversarially. The class of adversaries we consider includes as special cases those that invert or randomise their gradient estimate. On the practical side, we built our distributed training system in Pytorch. Benchmarking against the state of the art collective communications library (NCCL), our framework---with the parameter server housed entirely on one machine---led to a 25% reduction in time for training resnet50 on Imagenet when using 15 AWS p3.2xlarge machines.