Large-batch training has become a commonly used technique when training neural networks with a large number of GPU/TPU processors. As batch size increases, stochastic optimizers tend to converge to sharp local minima, leading to degraded test performance. Current methods usually use extensive data augmentation to increase the batch size, but we found the performance gain with data augmentation decreases as batch size increases, and data augmentation will become insufficient after certain point. In this paper, we propose to use adversarial learning to increase the batch size in large-batch training. Despite being a natural choice for smoothing the decision surface and biasing towards a flat region, adversarial learning has not been successfully applied in large-batch training since it requires at least two sequential gradient computations at each step, which will at least double the running time compared with vanilla training even with a large number of processors. To overcome this issue, we propose a novel Concurrent Adversarial Learning (ConAdv) method that decouple the sequential gradient computations in adversarial learning by utilizing staled parameters. Experimental results demonstrate that ConAdv can successfully increase the batch size on both ResNet-50 and EfficientNet training on ImageNet while maintaining high accuracy. In particular, we show ConAdv along can achieve 75.3\% top-1 accuracy on ImageNet ResNet-50 training with 96K batch size, and the accuracy can be further improved to 76.2\% when combining ConAdv with data augmentation. This is the first work successfully scales ResNet-50 training batch size to 96K.

Data parallelism does a good job in speeding up the training. However, when it comes to the case when the memory of a single device can not host a whole model, data parallelism would not have the chance to do anything. Another option is to split the model by operator, or horizontally. Megatron-LM introduced a 1-Dimensional distributed method to use GPUs to speed up the training process. Optimus is a 2D solution for distributed tensor parallelism. However, these methods have a high communication overhead and a low scaling efficiency on large-scale computing clusters. To solve this problem, we investigate the 2.5-Dimensional distributed tensor parallelism.Introduced by Solomonik et al., 2.5-Dimensional Matrix Multiplication developed an effective method to perform multiple Cannon's algorithm at the same time to increase the efficiency. With many restrictions of Cannon's Algorithm and a huge amount of shift operation, we need to invent a new method of 2.5-dimensional matrix multiplication to enhance the performance. Absorbing the essence from both SUMMA and 2.5-Dimensional Matrix Multiplication, we introduced SUMMA2.5-LM for language models to overcome the abundance of unnecessary transmission loss result from the increasing size of language model parallelism. Compared to previous 1D and 2D model parallelization of language models, our SUMMA2.5-LM managed to reduce the transmission cost on each layer, which could get a 1.45X efficiency according to our weak scaling result between 2.5-D [4,4,4] arrangement and 2-D [8,8,1] arrangement.

Many optimizers have been proposed for training deep neural networks, and they often have multiple hyperparameters, which make it tricky to benchmark their performance. In this work, we propose a new benchmarking protocol to evaluate both end-to-end efficiency (training a model from scratch without knowing the best hyperparameter) and data-addition training efficiency (the previously selected hyperparameters are used for periodically re-training the model with newly collected data). For end-to-end efficiency, unlike previous work that assumes random hyperparameter tuning, which over-emphasizes the tuning time, we propose to evaluate with a bandit hyperparameter tuning strategy. A human study is conducted to show that our evaluation protocol matches human tuning behavior better than the random search. For data-addition training, we propose a new protocol for assessing the hyperparameter sensitivity to data shift. We then apply the proposed benchmarking framework to 7 optimizers and various tasks, including computer vision, natural language processing, reinforcement learning, and graph mining. Our results show that there is no clear winner across all the tasks.

Large-batch training is an efficient approach for current distributed deep learning systems. It has enabled researchers to reduce the ImageNet/ResNet-50 training from 29 hours to around 1 minute. In this paper, we focus on studying the limit of the batch size. We think it may provide a guidance to AI supercomputer and algorithm designers. We provide detailed numerical optimization instructions for step-by-step comparison. Moreover, it is important to understand the generalization and optimization performance of huge batch training. Hoffer et al. introduced "ultra-slow diffusion" theory to large-batch training. However, our experiments show contradictory results with the conclusion of Hoffer et al. We provide comprehensive experimental results and detailed analysis to study the limitations of batch size scaling and "ultra-slow diffusion" theory. For the first time we scale the batch size on ImageNet to at least a magnitude larger than all previous work, and provide detailed studies on the performance of many state-of-the-art optimization schemes under this setting. We propose an optimization recipe that is able to improve the top-1 test accuracy by 18% compared to the baseline.

Large-batch training is key to speeding up deep neural network training in large distributed systems. However, large-batch training is difficult because it produces a generalization gap. Straightforward optimization often leads to accuracy loss on the test set. BERT \cite{devlin2018bert} is a state-of-the-art deep learning model that builds on top of deep bidirectional transformers for language understanding. Previous large-batch training techniques do not perform well for BERT when we scale the batch size (e.g. beyond 8192). BERT pre-training also takes a long time to finish (around three days on 16 TPUv3 chips). To solve this problem, we propose the LAMB optimizer, which helps us to scale the batch size to 65536 without losing accuracy. LAMB is a general optimizer that works for both small and large batch sizes and does not need hyper-parameter tuning besides the learning rate. The baseline BERT-Large model needs 1 million iterations to finish pre-training, while LAMB with batch size 65536/32768 only needs 8599 iterations. We push the batch size to the memory limit of a TPUv3 pod and can finish BERT training in 76 minutes.

Deep reinforcement learning (DRL) has gained a lot of attention in recent years, and has been proven to be able to play Atari games and Go at or above human levels. However, those games are assumed to have a small fixed number of actions and could be trained with a simple CNN network. In this paper, we study a special class of Asian popular card games called Dou Di Zhu, in which two adversarial groups of agents must consider numerous card combinations at each time step, leading to huge number of actions. We propose a novel method to handle combinatorial actions, which we call combinational Q-learning (CQL). We employ a two-stage network to reduce action space and also leverage order-invariant max-pooling operations to extract relationships between primitive actions. Results show that our method prevails over state-of-the art methods like naive Q-learning and A3C. We develop an easy-to-use card game environments and train all agents adversarially from sractch, with only knowledge of game rules and verify that our agents are comparative to humans. Our code to reproduce all reported results will be available online.

Large-batch training approaches have enabled researchers to utilize large-scale distributed processing and greatly accelerate deep-neural net (DNN) training. For example, by scaling the batch size from 256 to 32K, researchers have been able to reduce the training time of ResNet50 on ImageNet from 29 hours to 2.2 minutes (Ying et al., 2018). In this paper, we propose a new approach called linear-epoch gradual-warmup (LEGW) for better large-batch training. With LEGW, we are able to conduct large-batch training for both CNNs and RNNs with the Sqrt Scaling scheme. LEGW enables Sqrt Scaling scheme to be useful in practice and as a result we achieve much better results than the Linear Scaling learning rate scheme. For LSTM applications, we are able to scale the batch size by a factor of 64 without losing accuracy and without tuning the hyper-parameters. For CNN applications, LEGW is able to achieve the same accuracy even as we scale the batch size to 32K. LEGW works better than previous large-batch auto-tuning techniques. LEGW achieves a 5.3X average speedup over the baselines for four LSTM-based applications on the same hardware. We also provide some theoretical explanations for LEGW.

In this paper, we propose and study an Asynchronous parallel Greedy Coordinate Descent (Asy-GCD) algorithm for minimizing a smooth function with bounded constraints. At each iteration, workers asynchronously conduct greedy coordinate descent updates on a block of variables. In the first part of the paper, we analyze the theoretical behavior of Asy-GCD and prove a linear convergence rate. In the second part, we develop an efficient kernel SVM solver based on Asy-GCD in the shared memory multi-core setting. Since our algorithm is fully asynchronous--each core does not need to idle and wait for the other cores--the resulting algorithm enjoys good speedup and outperforms existing multi-core kernel SVM solvers including asynchronous stochastic coordinate descent and multi-core LIBSVM.