Global-QSGD accelerates distributed training by up to 3.51%.
problem High communication overhead in distributed deep learning.
method Allreduce-compatible gradient quantization with theoretical guarantees.
result Global-QSGD accelerates distributed training by up to 3.51%.
Pipe-SGD improves deep net training time by 5.4x with pipelined AllReduce.
problem Efficiently training deep neural networks on distributed systems.
method Pipelined SGD with two worker nodes, considering network latency and bandwidth.
result Up to 5.4x improvement in wall-clock training time.
Most commonly used distributed machine learning systems are either synchronous or centralized asynchronous. Synchronous algorithms like AllReduce-SGD perform poorly in a heterogeneous environment, while asynchronous algorithms using a parameter server suffer from 1) communication bottleneck at parameter servers when wo…
Mixed-precision CA-SGD for generalized linear models on GPUs
problem SGD communication bottleneck
method Mixed-precision CA-SGD
result Matches FP32 SGD loss within 0.5% on various problems
Applying machine learning techniques to the quickly growing data in science and industry requires highly-scalable algorithms. Large datasets are most commonly processed "data parallel" distributed across many nodes. Each node's contribution to the overall gradient is summed using a global allreduce. This allreduce is t…
End-to-end speech recognition system trained on GPUs and CPUs.
problem Building state-of-the-art speech recognition systems.
method Utilizes CPUs and GPUs for training, data augmentation, and neural network updates. Uses vocal tract length perturbation and acoustic simulator for data augmentation. Employed Horovod allreduce for training.
result Achieved 7.92% WER on proprietary English Bixby open domain test set using a Bidirectional Full Attention (BFA) model.
We present a system that enables rapid model experimentation for tera-scale machine learning with trillions of non-zero features, billions of training examples, and millions of parameters. Our contribution to the literature is a new method (SA L-BFGS) for changing batch L-BFGS to perform in near real-time by using stat…
SGP combines PushSum with stochastic gradient updates for robust distributed deep learning.
problem Synchronization issues in distributed deep learning.
method Stochastic Gradient Push (SGP) using PushSum for approximate distributed averaging.
result SGP converges to a stationary point at the same rate as SGD and achieves consensus.
Stochastic gradient descent (SGD) is a well known method for regression and classification tasks. However, it is an inherently sequential algorithm at each step, the processing of the current example depends on the parameters learned from the previous examples. Prior approaches to parallelizing linear learners using SG…
New algorithm improves distributed SGD with random sparsification for better convergence and generalization.
problem Communication bottleneck in distributed deep learning.
method Proposes detached error feedback (DEF) algorithm to improve convergence and generalization of communication-efficient distributed SGD.
result Shows better convergence and generalization bounds than existing methods.
SlowMo improves distributed SGD by periodically synchronizing and using momentum, leading to better accuracy.
problem Reducing communication overhead in distributed training while maintaining model accuracy.
method Periodically synchronizing workers and using momentum after multiple iterations of a base optimization algorithm.
result SlowMo consistently yields improvements in optimization and generalization performance relative to the base optimizer.
New algorithm reduces communication bandwidth for large-scale deep learning training.
problem Efficiently compressing gradients for ring all-reduce in large-scale clusters.
method Importance weighted pruning based on gradient and parameter size.
result Achieved significant gradient compression ratios (64X and 58.8X) on AlexNet and ResNet50.
NEST optimizes deep learning training by placing devices efficiently across networks and memory.
problem Inefficient device placement in distributed deep learning leads to high communication and memory overhead.
method NEST uses network-, compute-, and memory-aware dynamic programming to optimize device placement.
result NEST achieves up to 2.43 times higher throughput and better memory efficiency.
ADPSGD enables faster ASR training with larger batch sizes.
problem Efficient distributed deep learning for ASR with large batch sizes.
method Asynchronous Decentralized Parallel Stochastic Gradient Descent (ADPSGD) and Hierarchical-ADPSGD (H-ADPSGD).
result ADPSGD can converge with a 3X larger batch size than SSGD, enabling faster training.
Mesh-TensorFlow enables efficient deep learning on large clusters.
problem Memory constraints and inefficiency in batch-splitting for large models.
method Introduces Mesh-TensorFlow for specifying general tensor computations across a multi-dimensional mesh of processors.
result Trains Transformer models with up to 5 billion parameters on TPU meshes of up to 512 cores.
CodedReduce combines tree topology and gradient coding for efficient and resilient gradient aggregation.
problem Efficient and robust gradient aggregation in distributed learning.
method CodedReduce combines tree topology and gradient coding to overcome bandwidth bottlenecks and straggler delays.
result CodedReduce achieves up to 27.2x speedup over benchmarks GC and RAR.