Deep learning models trained on large data sets have been widely successful in both vision and language domains. As state-of-the-art deep learning architectures have continued to grow in parameter count so have the compute budgets and times required to train them, increasing the need for compute-efficient methods that parallelize training. Two common approaches to parallelize the training of deep networks have been data and model parallelism. While useful, data and model parallelism suffer from diminishing returns in terms of compute efficiency for large batch sizes. In this paper, we investigate how to continue scaling compute efficiently beyond the point of diminishing returns for large batches through local parallelism, a framework which parallelizes training of individual layers in deep networks by replacing global backpropagation with truncated layer-wise backpropagation. Local parallelism enables fully asynchronous layer-wise parallelism with a low memory footprint, and requires little communication overhead compared with model parallelism. We show results in both vision and language domains across a diverse set of architectures, and find that local parallelism is particularly effective in the high-compute regime.
In data-parallel synchronous training of deep neural networks, different devices (replicas) run the same program with different partitions of the training batch, but weight update computation is repeated on all replicas, because the weights do not have a batch dimension to partition. This can be a bottleneck for performance and scalability in typical language models with large weights, and models with small per-replica batch size which is typical in large-scale training. This paper presents an approach to automatically shard the weight update computation across replicas with efficient communic…
Mean-Field Parallel Decoding for Discrete Diffusion Language Models
School of Electrical & Computer Engineering, Tel Aviv University（特拉维夫大学电气与计算机工程学院）；School of Computer Science and AI, Tel Aviv University（特拉维夫大学计算机科学与人工智能学院）；Department of Computer Science, Technion, Israel Institute of Technology（以色列理工学院计算机科学系）
University of Turin（都灵大学）；Synesthesia s.r.l.（联觉有限公司）；Télécom Paris（巴黎电信学院）；Institut Polytechnique de Paris（巴黎理工学院）
Recent advances in deep learning optimization showed that, with some a-posteriori information on fully-trained models, it is possible to match the same performance by simply training a subset of their parameters. Such a discovery has a broad impact from theory to applications, driving the research towards methods to identify the minimum subset of parameters to train without look-ahead information exploitation. However, the methods proposed do not match the state-of-the-art performance, and rely on unstructured sparsely connected models. In this work we shift our focus from the single parameter…
Effective regularisation during training can mean the difference between success and failure for deep neural networks. Recently, dither has been suggested as alternative to dropout for regularisation during batch-averaged stochastic gradient descent (SGD). In this article, we show that these methods fail without batch averaging and we introduce a new, parallel regularisation method that may be used without batch averaging. Our results for parallel-regularised non-batch-SGD are substantially better than what is possible with batch-SGD. Furthermore, our results demonstrate that dither and dropou…
Faculty of Informatics（信息学院）；Universià della Svizzera italiana（瑞士意大利语区大学）
The creation of practical deep learning data-products often requires parallelization across processors and computers to make deep learning feasible on large data sets, but bottlenecks in communication bandwidth make it difficult to attain good speedups through parallelism. Here we develop and test 8-bit approximation algorithms which make better use of the available bandwidth by compressing 32-bit gradients and nonlinear activations to 8-bit approximations. We show that these approximations do not decrease predictive performance on MNIST, CIFAR10, and ImageNet for both model and data paralleli…
We build on the recently proposed EigenGame that views eigendecomposition as a competitive game. EigenGame's updates are biased if computed using minibatches of data, which hinders convergence and more sophisticated parallelism in the stochastic setting. In this work, we propose an unbiased stochastic update that is asymptotically equivalent to EigenGame, enjoys greater parallelism allowing computation on datasets of larger sample sizes, and outperforms EigenGame in experiments. We present applications to finding the principal components of massive datasets and performing spectral clustering o…
Deep Neural Networks (DNNs) are becoming an important tool in modern computing applications. Accelerating their training is a major challenge and techniques range from distributed algorithms to low-level circuit design. In this survey, we describe the problem from a theoretical perspective, followed by approaches for its parallelization. We present trends in DNN architectures and the resulting implications on parallelization strategies. We then review and model the different types of concurrency in DNNs: from the single operator, through parallelism in network inference and training, to distri…
Deep Learning (DL) models are becoming larger, because the increase in model size might offer significant accuracy gain. To enable the training of large deep networks, data parallelism and model parallelism are two well-known approaches for parallel training. However, data parallelism does not help reduce memory footprint per device. In this work, we introduce Large deep 3D ConvNets with Automated Model Parallelism (LAMP) and investigate the impact of both input's and deep 3D ConvNets' size on segmentation accuracy. Through automated model parallelism, it is feasible to train large deep 3D Con…
We present a technique, which we term leapfrogging, to parallelize back- propagation in deep neural networks. We show that this technique yields a savings of $1-1/k$ of a dominant term in backpropagation, where k is the number of threads (or gpus).
The Technion - Israel Institute of Technology（以色列理工学院）；UC Berkeley（加州大学伯克利分校）
Deep reinforcement learning (DRL) methods such as the Deep Q-Network (DQN) have achieved state-of-the-art results in a variety of challenging, high-dimensional domains. This success is mainly attributed to the power of deep neural networks to learn rich domain representations for approximating the value function or policy. Batch reinforcement learning methods with linear representations, on the other hand, are more stable and require less hyper parameter tuning. Yet, substantial feature engineering is necessary to achieve good results. In this work we propose a hybrid approach -- the Least Squ…
School of Computer Science Carnegie Mellon University（卡内基梅隆大学计算机科学学院）
When training large machine learning models with many variables or parameters, a single machine is often inadequate since the model may be too large to fit in memory, while training can take a long time even with stochastic updates. A natural recourse is to turn to distributed cluster computing, in order to harness additional memory and processors. However, naive, unstructured parallelization of ML algorithms can make inefficient use of distributed memory, while failing to obtain proportional convergence speedups - or can even result in divergence. We develop a framework of primitives for dyna…
We show that deep networks can be trained using Hebbian updates yielding similar performance to ordinary back-propagation on challenging image datasets. To overcome the unrealistic symmetry in connections between layers, implicit in back-propagation, the feedback weights are separate from the feedforward weights. The feedback weights are also updated with a local rule, the same as the feedforward weights - a weight is updated solely based on the product of activity of the units it connects. With fixed feedback weights as proposed in Lillicrap et. al (2016) performance degrades quickly as the d…
Although deep neural networks (DNN) are able to scale with direct advances in computational power (e.g., memory and processing speed), they are not well suited to exploit the recent trends for parallel architectures. In particular, gradient descent is a sequential process and the resulting serial dependencies mean that DNN training cannot be parallelized effectively. Here, we show that a DNN may be replicated over a massive parallel architecture and used to provide a cumulative sampling of local solution space which results in rapid and robust learning. We introduce a complimentary convolution…
Nara Institute of Science and Technology（奈良先端科学技术大学院大学）
This paper proposes a new robust update rule of target network for deep reinforcement learning (DRL), to replace the conventional update rule, given as an exponential moving average. The target network is for smoothly generating the reference signals for a main network in DRL, thereby reducing learning variance. The problem with its conventional update rule is the fact that all the parameters are smoothly copied with the same speed from the main network, even when some of them are trying to update toward the wrong directions. This behavior increases the risk of generating the wrong reference s…
Yonsei University（延世大学）；University of California at Los Angeles (UCLA)（加州大学洛杉矶分校（UCLA））
In this paper, we propose a novel model-parallel learning method, called local critic training, which trains neural networks using additional modules called local critic networks. The main network is divided into several layer groups and each layer group is updated through error gradients estimated by the corresponding local critic network. We show that the proposed approach successfully decouples the update process of the layer groups for both convolutional neural networks (CNNs) and recurrent neural networks (RNNs). In addition, we demonstrate that the proposed method is guaranteed to conver…
Deep learning systems have become vital tools across many fields, but the increasing model sizes mean that training must be accelerated to maintain such systems' utility. Current systems like Tensorflow and MXNet focus on one specific parallelization strategy, data parallelism, which requires large training batch sizes in order to scale. We cast the problem of finding the best parallelization strategy as the problem of finding the best tiling to partition tensors with the least overall communication. We propose an algorithm that can find the optimal tiling. Our resulting parallelization soluti…
Training deep neural networks (DNNs) in large-cluster computing environments is increasingly necessary, as networks grow in size and complexity. Local memory and processing limitations require robust data and model parallelism for crossing compute node boundaries. We propose a linear-algebraic approach to model parallelism in deep learning, which allows parallel distribution of any tensor in the DNN. Rather than rely on automatic differentiation tools, which do not universally support distributed memory parallelism models, we show that parallel data movement operations, e.g., broadcast, sum-re…
We propose Episodic Backward Update (EBU) - a novel deep reinforcement learning algorithm with a direct value propagation. In contrast to the conventional use of the experience replay with uniform random sampling, our agent samples a whole episode and successively propagates the value of a state to its previous states. Our computationally efficient recursive algorithm allows sparse and delayed rewards to propagate directly through all transitions of the sampled episode. We theoretically prove the convergence of the EBU method and experimentally demonstrate its performance in both deterministic…
Rectified Linear Units (ReLU) seem to have displaced traditional 'smooth' nonlinearities as activation-function-du-jour in many - but not all - deep neural network (DNN) applications. However, nobody seems to know why. In this article, we argue that ReLU are useful because they are ideal demodulators - this helps them perform fast abstract learning. However, this fast learning comes at the expense of serious nonlinear distortion products - decoy features. We show that Parallel Dither acts to suppress the decoy features, preventing overfitting and leaving the true features cleanly demodulated f…
University of Torino（都灵大学）；Noesis Solutions NV（Noesis Solutions NV公司）；Pacific Northwest National Lab.（太平洋西北国家实验室）
This work aims to assess the state of the art of data parallel deep neural network training, trying to identify potential research tracks to be exploited for performance improvement. Beside, it presents a design for a practical C++ library dedicated at implementing and unifying the current state of the art methodologies for parallel training in a performance-conscious framework, allowing the user to explore novel strategies without departing significantly from its usual work-flow.
Training large machine learning (ML) models with many variables or parameters can take a long time if one employs sequential procedures even with stochastic updates. A natural solution is to turn to distributed computing on a cluster; however, naive, unstructured parallelization of ML algorithms does not usually lead to a proportional speedup and can even result in divergence, because dependencies between model elements can attenuate the computational gains from parallelization and compromise correctness of inference. Recent efforts toward this issue have benefited from exploiting the static, …
As deep learning becomes more expensive, both in terms of time and compute, inefficiencies in machine learning (ML) training prevent practical usage of state-of-the-art models for most users. The newest model architectures are simply too large to be fit onto a single processor. To address the issue, many ML practitioners have turned to model parallelism as a method of distributing the computational requirements across several devices. Unfortunately, the sequential nature of neural networks causes very low efficiency and device utilization in model parallel training jobs. We propose a new form …
The most popular framework for distributed training of machine learning models is the (synchronous) parameter server (PS). This paradigm consists of $n$ workers, which iteratively compute updates of the model parameters, and a stateful PS, which waits and aggregates all updates to generate a new estimate of model parameters and sends it back to the workers for a new iteration. Transient computation slowdowns or transmission delays can intolerably lengthen the time of each iteration. An efficient way to mitigate this problem is to let the PS wait only for the fastest $n-b$ updates, before gener…
Massachusetts Institute of Technology Lincoln Laboratory Supercomputing Center（麻省理工学院林肯实验室超级计算中心）
A Multigrid Full Approximation Storage algorithm for solving Deep Residual Networks is developed to enable neural network parallelized layer-wise training and concurrent computational kernel execution on GPUs. This work demonstrates a 10.2x speedup over traditional layer-wise model parallelism techniques using the same number of compute units.
Institute of Information Science, Academia Sinica（中央研究院信息科学研究所）；Research Center for Information Technology Innovation, Academia Sinica（中央研究院信息技术创新研究中心）
In this paper, we propose a dictionary update method for Nonnegative Matrix Factorization (NMF) with high dimensional data in a spectral conversion (SC) task. Voice conversion has been widely studied due to its potential applications such as personalized speech synthesis and speech enhancement. Exemplar-based NMF (ENMF) emerges as an effective and probably the simplest choice among all techniques for SC, as long as a source-target parallel speech corpus is given. ENMF-based SC systems usually need a large amount of bases (exemplars) to ensure the quality of the converted speech. However, a sma…
Using multiple nodes and parallel computing algorithms has become a principal tool to improve training and execution times of deep neural networks as well as effective collective intelligence in sensor networks. In this paper, we consider the parallel implementation of an already-trained deep model on multiple processing nodes (a.k.a. workers) where the deep model is divided into several parallel sub-models, each of which is executed by a worker. Since latency due to synchronization and data transfer among workers negatively impacts the performance of the parallel implementation, it is desirab…
The recent success of deep learning applications has coincided with those widely available powerful computational resources for training sophisticated machine learning models with huge datasets. Nonetheless, training large models such as convolutional neural networks using model parallelism (as opposed to data parallelism) is challenging because the complex nature of communication between model shards makes it difficult to partition the computation efficiently across multiple machines with an acceptable trade-off. This paper presents SplitBrain, a high performance distributed deep learning fra…
Training deep neural networks is a challenging non-convex optimization problem. Recent work has proven that the strong duality holds (which means zero duality gap) for regularized finite-width two-layer ReLU networks and consequently provided an equivalent convex training problem. However, extending this result to deeper networks remains to be an open problem. In this paper, we prove that the duality gap for deeper linear networks with vector outputs is non-zero. In contrast, we show that the zero duality gap can be obtained by stacking standard deep networks in parallel, which we call a paral…
Measurement update rules for Bayes filters often contain hand-crafted heuristics to compute observation probabilities for high-dimensional sensor data, like images. In this work, we propose the novel approach Deep Measurement Update (DMU) as a general update rule for a wide range of systems. DMU has a conditional encoder-decoder neural network structure to process depth images as raw inputs. Even though the network is trained only on synthetic data, the model shows good performance at evaluation time on real-world data. With our proposed training scheme primed data training , we demonstrate ho…
AMDP: Asynchronous Multi-Directional Pipeline Parallelism for Large-Scale Models Training
State Key Laboratory of Blockchain and Data Security, Zhejiang University, Hangzhou, China（区块链与数据安全国家重点实验室，浙江大学，杭州，中国）；College of Computer Science and Technology, Zhejiang University, Hangzhou, China（计算机科学与技术学院，浙江大学，杭州，中国）
UC Berkeley（加州大学伯克利分校）；Amazon Web Services（亚马逊云科技）；Google（谷歌）；Shanghai Jiao Tong University（上海交通大学）；MBZUAI（穆罕默德·本·扎耶德人工智能大学）；Carnegie Mellon University（卡内基梅隆大学）；Duke University（杜克大学）
Alpa automates model-parallel training of large deep learning (DL) models by generating execution plans that unify data, operator, and pipeline parallelism. Existing model-parallel training systems either require users to manually create a parallelization plan or automatically generate one from a limited space of model parallelism configurations. They do not suffice to scale out complex DL models on distributed compute devices. Alpa distributes the training of large DL models by viewing parallelisms as two hierarchical levels: inter-operator and intra-operator parallelisms. Based on it, Alpa c…
