Parallel Restarted SGD with Faster Convergence and Less Communication: Demystifying Why Model Averaging Works for Deep Learning

Neural network training is inherently sequential where the layers finish the forward propagation in succession, followed by the calculation and back-propagation of gradients (based on a loss function) starting from the last layer. The sequential computations significantly slow down neural network training, especially the deeper ones. Prediction has been successfully used in many areas of computer architecture to speed up sequential processing. Therefore, we propose ADA-GP, which uses gradient prediction adaptively to speed up deep neural network (DNN) training while maintaining accuracy. ADA-GP works by incorporating a small neural network to predict gradients for different layers of a DNN model. ADA-GP uses a novel tensor reorganization method to make it feasible to predict a large number of gradients. ADA-GP alternates between DNN training using backpropagated gradients and DNN training using predicted gradients. ADA-GP adaptively adjusts when and for how long gradient prediction is used to strike a balance between accuracy and performance. Last but not least, we provide a detailed hardware extension in a typical DNN accelerator to realize the speed up potential from gradient prediction. Our extensive experiments with fifteen DNN models show that ADA-GP can achieve an average speed up of 1.47 × with similar or even higher accuracy than the baseline models. Moreover, it consumes, on average, 34% less energy due to reduced off-chip memory accesses compared to the baseline accelerator.

A variety of methods have been applied to the architectural configuration and learning or training of artificial deep neural networks (DNN). These methods play a crucial role in the success or failure of the DNN for most problems and applications. Evolutionary algorithms (EAs) are gaining momentum as a computationally feasible method for the automated optimization of DNNs. Neuroevolution is a term, which describes these processes of automated configuration and training of DNNs using EAs. While many works exist in the literature, no comprehensive surveys currently exist focusing exclusively on the strengths and limitations of using neuroevolution approaches in DNNs. Absence of such surveys can lead to a disjointed and fragmented field preventing DNNs researchers potentially adopting neuroevolutionary methods in their own research, resulting in lost opportunities for wider application within real-world deep learning problems. This article presents a comprehensive survey, discussion, and evaluation of the state-of-the-art in using EAs for architectural configuration and training of DNNs. This article highlights the most pertinent current issues and challenges in neuroevolution and identifies multiple promising future research directions. <p xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><i>Impact Statement—</i>The concept of deep learning originated from the study of artificial neural networks (ANNs). ANNs have achieved extraordinary results in a variety of diverse application areas. Numerous methods have been applied to the architectural configuration and learning or training of artificial DNN and these methods play a crucial role in the success or failure of the DNN for most problems and applications. Recently, EAs have been gaining momentum as a computationally feasible method (called neuroevolution) for the automated configuration and learning or training of DNNs. This article reviews over 170 recent scientific papers describing how major EAs paradigms are being applied by researchers to the configuration and optimization of multiple DNNs. By articulating a clear understanding of the context, state-of-the-art, and feasibility of Neuroevolution, researchers in AI, EAs, and DNN will benefit from this article. The impact of this article comes from contributing toward enhancing research capacity, knowledge, and skills for researchers currently working in neuroevolution and actively engaging those considering becoming involved in this area.

First-come first-serve scheduling can result in substantial (up to 10%) of transiently idle nodes on supercomputers. Recognizing that such unfilled nodes are well-suited for deep neural network (DNN) training, due to the flexible nature of DNN training tasks, Liu et al. proposed that the re-scaling DNN training tasks to fit gaps in schedules be formulated as a mixed-integer linear programming (MILP) problem, and demonstrated via simulation the potential benefits of the approach. Here, we introduce MalleTrain, a system that provides the first practical implementation of this approach and that furthermore generalizes it by allowing it to be used even for DNN training applications for which model information is unknown before runtime. Key to this latter innovation is the use of a lightweight online job profiling advisor (JPA) to collect critical scalability information for DNN jobs---information that it then employs to optimize resource allocations dynamically, in real time. We describe the MalleTrain architecture and present the results of a detailed experimental evaluation on a supercomputer GPU cluster and several representative DNN training workloads, including neural architecture search and hyperparameter optimization. Our results not only confirm the practical feasibility of leveraging idle supercomputer nodes for DNN training but improve significantly on prior results, improving training throughput by up to 22.3% without requiring users to provide job scalability information.

Distributed training in deep learning (DL) is common practice as data and models grow. The current practice for distributed training of deep neural networks faces the challenges of communication bottlenecks when operating at scale, and model accuracy deterioration with an increase in global batch size. Present solutions focus on improving message exchange efficiency as well as implementing techniques to tweak batch sizes and models in the training process. The loss of training accuracy typically happens because the loss function gets trapped in a local minima. We observe that the loss landscape minimization is shaped by both the model and training data and propose a data optimization approach that utilizes machine learning to implicitly smooth out the loss landscape resulting in fewer local minima. Our approach filters out data points which are less important to feature learning, enabling us to speed up the training of models on larger batch sizes to improved accuracy.

Deep Learning (DL) has become a prominent machine learning technique due to the availability of efficient computational resources in the form of Graphics Processing Units (GPUs), large-scale datasets and a variety of models. The newer generation of GPUs are being designed with special emphasis on optimizing performance for DL applications. Also, the availability of easy-to-use DL frameworks—like PyTorch and TensorFlow— has enhanced productivity of domain experts to work on their custom DL applications from diverse domains. However, existing Deep Neural Network (DNN) training approaches may not fully utilize the newly emerging powerful GPUs like the NVIDIA A100—this is the primary issue that we address in this paper. Our motivating analyses show that the GPU utilization on NVIDIA A100 can be as low as 43% using traditional DNN training approaches for small-to-medium DL models and input data size. This paper proposes AccDP—a data-parallel distributed DNN training approach—to accelerate GPU-based DL applications. AccDP exploits the Message Passing Interface (MPI) communication library coupled with the NVIDIA’s Multi-Process Service (MPS) to increase the amount of work assigned to parallel GPUs resulting in higher utilization of compute resources. We evaluate our proposed design on different small-to-medium DL models and input sizes on the state-of-the-art HPC clusters. By injecting more parallelism into DNN training using our approach, the evaluation shows up to 58% improvement in training performance on a single GPU and up to 62% on 16 GPUs compared to regular DNN training. Furthermore, we conduct an in-depth characterization to determine the impact of several DNN training factors and best practices—including the batch size and the number of data loading workers— to optimally utilize GPU devices. To the best of our knowledge, this is the first work that explores the use of MPS and MPI to maximize the utilization of GPUs in distributed DNN training.

Deep neural network (DNN) training is generally performed by cloud computing platforms. However, cloud-based training has several problems such as network bottleneck, server management cost, and privacy. To overcome these problems, one of the most promising solutions is distributed DNN model training which trains the model with not only high-performance servers but also low-end power-efficient mobile edge or user devices. However, due to the lack of a framework which can provide an optimal cluster configuration (i.e., determining which computing devices participate in DNN training tasks), it is difficult to perform efficient DNN model training considering DNN service providers' preferences such as training time or energy efficiency. In this paper, we introduce a novel framework for distributed DNN training that determines the best training cluster configuration with available heterogeneous computing resources. Our proposed framework utilizes pre-training with a small number of training steps and estimates training time, power, energy, and energy-delay product (EDP) for each possible training cluster configuration. Based on the estimated metrics, our framework performs DNN training for the remaining steps with the chosen best cluster configurations depending on DNN service providers' preferences. Our framework is implemented in TensorFlow and evaluated with three heterogeneous computing platforms and five widely used DNN models. According to our experimental results, in 76.67% of the cases, our framework chooses the best cluster configuration depending on DNN service providers' preferences with only a small training time overhead.

Learning Rate (LR) is an important hyper-parameter to tune for effective training of deep neural networks (DNNs). Even for the baseline of a constant learning rate, it is non-trivial to choose a good constant value for training a DNN. Dynamic learning rates involve multi-step tuning of LR values at various stages of the training process and offer high accuracy and fast convergence. However, they are much harder to tune. In this paper, we present a comprehensive study of 13 learning rate functions and their associated LR policies by examining their range parameters, step parameters, and value update parameters. We propose a set of metrics for evaluating and selecting LR policies, including the classification confidence, variance, cost, and robustness, and implement them in LRBench, an LR benchmarking system. LRBench can assist end-users and DNN developers to select good LR policies and avoid bad LR policies for training their DNNs. We tested LRBench on Caffe, an open source deep learning framework, to showcase the tuning optimization of LR policies. Evaluated through extensive experiments, we attempt to demystify the tuning of LR policies by identifying good LR policies with effective LR value ranges and step sizes for LR update schedules.

EC-DNN: A new method for parallel training of deep neural networks

Parallelization framework has become a necessity to speed up the training of deep neural networks (DNN) recently. Such framework typically employs the Model Average approach, denoted as MA-DNN, in which parallel workers conduct respective training based on their own local data while the parameters of local models are periodically communicated and averaged to obtain a global model which serves as the new start of local models. However, since DNN is a highly non-convex model, averaging parameters cannot ensure that such global model can perform better than those local models. To tackle this problem, we introduce a new parallel training framework called Ensemble-Compression, denoted as EC-DNN. In this framework, we propose to aggregate the local models by ensemble, i.e., averaging the outputs of local models instead of the parameters. As most of prevalent loss functions are convex to the output of DNN, the performance of ensemble-based global model is guaranteed to be at least as good as the average performance of local models. However, a big challenge lies in the explosion of model size since each round of ensemble can give rise to multiple times size increment. Thus, we carry out model compression after each ensemble, specialized by a distillation based method in this paper, to reduce the size of the global model to be the same as the local ones. Our experimental results demonstrate the prominent advantage of EC-DNN over MA-DNN in terms of both accuracy and speedup.

Recently deep-learning (DL) techniques have been widely adopted in side-channel power analysis. A DL-assisted SCA generally consists of two phases: a deep neural network (DNN) training phase and a follow-on attack phase using the trained DNN. However, currently the two phases are not well aligned, as there is no conclusion on what metric used in the training can result in the most effective attack in the second phase. When traditional loss functions such as negative log-likelihood (NLL) are used in training a DNN, the trained model does not yield optimal follow-on attack. Recently some information theoretical SCA leakage metrics are proposed, either as the validation metric to stop the DNN training with traditional loss functions, or as both the validation metric and the training loss function. None of those proposed metrics, however, directly measures the SCA effectiveness. We propose to conduct DNN training directly with a common SCA effectiveness metric, Guessing Entropy (GE). We overcome the prior practical difficulty of using GE in DNN training by utilizing the GEEA estimation algorithm introduced in CHES 2020. We show that using GEEA as either the validation metric or the loss function produces DNN models that lead to much more effective follow-on attacks. Our work consolidates the DL-assisted SCA framework with a consistent metric, which shows great potential to be adopted as the universal SCA-oriented DNN training framework.

Context-dependent deep neural network HMMs have been shown to achieve recognition accuracy superior to Gaussian mixture models in a number of recent works. Typically, neural networks are optimized with stochastic gradient descent. On large datasets, stochastic gradient descent improves quickly during the beginning of the optimization. But since it does not make use of second order information, its asymptotic convergence behavior is slow. In regions with pathological curvature, stochastic gradient descent may almost stagnate and thereby falsely indicate convergence. Another drawback of stochastic gradient descent is that it can only be parallelized within minibatches. The Hessian-free algorithm is a second order batch optimization algorithm that does not suffer from these problems. In a recent work, Hessian-free optimization has been applied to a training of deep neural networks according to a sequence criterion. In that work, improvements in accuracy and training time have been reported. In this paper, we analyze the properties of the Hessian-free optimization algorithm and investigate whether it is suited for cross-entropy training of deep neural networks as well.

Constructed encoded data based coded distributed DNN training for edge computing scenario

As the popularity of deep learning in industry rapidly grows, efficient training of deep neural networks (DNNs) becomes important. To train a DNN with a large amount of data, distributed training with data parallelism has been widely adopted. However, the communication overhead limits the scalability of distributed training. To reduce the overhead, a number of distributed training algorithms have been proposed. The model accuracy and training performance of those algorithms can be different depending on various factors such as cluster settings, training models/datasets, and optimization techniques applied. In order for someone to adopt a distributed training algorithm appropriate for her/his situation, it is required for her/him to fully understand the model accuracy and training performance of these algorithms in various settings. Toward this end, this paper reviews and evaluates seven popular distributed training algorithms (BSP, ASP, SSP, EASGD, AR-SGD, GoSGD, and AD-PSGD) in terms of the model accuracy and training performance in various settings. Specifically, we evaluate those algorithms for two CNN models, in different cluster settings, and with three well-known optimization techniques. Through extensive evaluation and analysis, we made several interesting discoveries. For example, we found out that some distributed training algorithms (SSP, EASGD, and GoSGD) have highly negative impact on the model accuracy because they adopt intermittent and asymmetric communication to improve training performance; the communication overhead of some centralized algorithms (ASP and SSP) is much higher than we expected in a cluster setting with limited network bandwidth because of the PS bottleneck problem. These findings, and many more in the paper, can guide the adoption of proper distributed training algorithms in industry; our findings can be useful in academia as well for designing new distributed training algorithms.

OptQuant: Distributed training of neural networks with optimized quantization mechanisms

Low-bit-width data formats offer a promising solution for enhancing the energy efficiency of Deep Neural Network (DNN) training accelerators. In this work, we introduce a novel 5.3-bit data format that groups fixed-point values sharing a common exponent and scaling factor within a block of data. We propose a two-level logarithmic mantissa scaling method, providing a wide dynamic range for the Block Floating Point based data format and mitigating data loss during the conversion of DNN parameters into the proposed format. Evaluation results demonstrate that the proposed data format enables DNN training with an average of 5.3 bits, with minimal accuracy loss and no required modifications to the training process. Additionally, implementation results of proposed data format shows 1.48<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\boldsymbol{\times}$</tex-math></inline-formula> reduction in memory footprint compared to 8-bit floating point formats, while addressing the lack of sufficient dynamic range for DNN training in current low-bit-width formats such as 8-bit and below.