Efficient Training of Convolutional Neural Nets on Large Distributed Systems

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.

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.

In this study we consider adaptive power beaming with a fiber-array laser transmitter system in presence of atmospheric turbulence. For optimization of power transition through the atmosphere a fiber-array is traditionally controlled by stochastic parallel gradient descent (SPGD) algorithm where control feedback is provided via a radio frequency link by an optical-to-electrical power conversion sensor, attached to a cooperative target. The SPGD algorithm continuously and randomly perturbs voltages applied to fiber-array phase shifters and fiber tip positioners in order to maximize sensor signal, i.e. uses, the so-called, “blind” optimization principle. By contrast to this approach a prospective artificially intelligent (AI) control systems for synthesis of optimal control can utilize various pupil- or target-plane data available for the analysis including wavefront sensor data, photo-voltaic array (PVA) data, other optical or atmospheric parameters, and potentially can eliminate well-known drawbacks of SPGDbased controllers. In this study an optimal control is synthesized by a deep neural network (DNN) using target-plane PVA sensor data as its input. A DNN training is occurred online in sync with control system operation and is performed by applying of small perturbations to DNN’s outputs. This approach does not require initial DNN’s pre-training as well as guarantees optimization of system performance in time. All theoretical results are verified by numerical experiments.

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.

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

Despite simplicity, stochastic gradient descent (SGD)-like algorithms are successful in training deep neural networks (DNNs). Among various attempts to improve SGD, weight averaging (WA), which averages the weights of multiple models, has recently received much attention in the literature. Broadly, WA falls into two categories: 1) online WA, which averages the weights of multiple models trained in parallel, is designed for reducing the gradient communication overhead of parallel mini-batch SGD and 2) offline WA, which averages the weights of one model at different checkpoints, is typically used to improve the generalization ability of DNNs. Though online and offline WA are similar in form, they are seldom associated with each other. Besides, these methods typically perform either offline parameter averaging or online parameter averaging, but not both. In this work, we first attempt to incorporate online and offline WA into a general training framework termed hierarchical WA (HWA). By leveraging both the online and offline averaging manners, HWA is able to achieve both faster convergence speed and superior generalization performance without any fancy learning rate adjustment. Besides, we also analyze the issues faced by the existing WA methods, and how our HWA addresses them, empirically. Finally, extensive experiments verify that HWA outperforms the state-of-the-art methods significantly.

We consider in this work a system of two stochastic differential equations named the perturbed compositional gradient flow. By introducing a separation of fast and slow scales of the two equations, we show that the limit of the slow motion is given by an averaged ordinary differential equation. We then demonstrate that the deviation of the slow motion from the averaged equation, after proper rescaling, converges to a stochastic process with Gaussian inputs. This indicates that the slow motion can be approximated in the weak sense by a standard perturbed gradient flow or the continuous-time stochastic gradient descent algorithm that solves the optimization problem for a composition of two functions. As an application, the perturbed compositional gradient flow corresponds to the diffusion limit of the Stochastic Composite Gradient Descent (SCGD) algorithm for minimizing a composition of two expected-value functions in the optimization literatures. For the strongly convex case, such an analysis implies that the SCGD algorithm has the same convergence time asymptotic as the classical stochastic gradient descent algorithm. Thus it validates, at the level of continuous approximation, the effectiveness of using the SCGD algorithm in the strongly convex case.

Linear prediction methods, such as least squares for regression, logistic regression and support vector machines for classification, have been extensively used in statistics and machine learning. In this paper, we study stochastic gradient descent (SGD) algorithms on regularized forms of linear prediction methods. This class of methods, related to online algorithms such as perceptron, are both efficient and very simple to implement. We obtain numerical rate of convergence for such algorithms, and discuss its implications. Experiments on text data will be provided to demonstrate numerical and statistical consequences of our theoretical findings.

In the paper, we consider a stochastic gradient descent algorithm in combination with a dropout method. We used the theory of intuitionistic fuzzy sets for the assessment of the equivalence of the respective assessment units. We also consider a degree of uncertainty when the information is not enough.

In this invited paper, we present deep neural network (DNN) training accelerator designs in both ASIC and FPGA. The accelerators implements stochastic gradient descent based training algorithm in 16-bit fixed-point precision. A new cyclic weight storage and access scheme enables using the same off-the-shelf SRAMs for non-transpose and transpose operations during feed-forward and feed-backward phases, respectively, of the DNN training process. Including the cyclic weight scheme, the overall DNN training processor is implemented in both 65nm CMOS ASIC and Intel Stratix-10 FPGA hardware. We collectively report the ASIC and FPGA training accelerator results.

Structural health monitoring (SHM) provides an economic approach which aims to enhance understanding the behavior of structures by continuously collecting data through multiple networked sensors attached to the structure. These data are then utilized to gain insight into the health of a structure and make timely and economic decisions about its maintenance. The generated SHM sensing data are non-stationary and exist in a correlated multi-way form which makes the batch/off-line learning and standard two-way matrix analysis unable to capture all of these correlations and relationships. In this sense, the online tensor data analysis has become an essential tool for capturing underlying structures in higher-order datasets stored in a tensor $${\mathcal {X}} \in {\mathbb {R}} ^{I_1 \times \cdots \times I_N} $$ X ∈ R I 1 × ⋯ × I N . The CANDECOMP/PARAFAC (CP) decomposition has been extensively studied and applied to approximate X by N loading matrices $$A(1),\ldots ,A(N)$$ A ( 1 ) , … , A ( N ) where N represents the order of the tensor. We propose a novel algorithm, FP-CPD, to parallelize the CANDECOMP/PARAFAC (CP) decomposition of a tensor $${\mathcal {X}} \in {\mathbb {R}} ^{I_1 \times \cdots \times I_N} $$ X ∈ R I 1 × ⋯ × I N . Our approach is based on stochastic gradient descent (SGD) algorithm which allows us to parallelize the learning process, and it is very useful in online setting since it updates $${\mathcal {X}}^{t+1}$$ X t + 1 in one single step. Our SGD algorithm is augmented with Nesterov’s accelerated gradient and perturbation methods to accelerate and guarantee convergence. The experimental results using laboratory-based and real-life structural datasets indicate fast convergence and good scalability.

Stochastic differential equation (SDE) models are the foundation for pricing and hedging financial derivatives. The drift and volatility functions in SDE models are typically chosen to be algebraic functions with a small number ( < 5 ) of parameters which can be calibrated to market data. A more flexible approach is to use neural networks to model the drift and volatility functions, which provides more degrees-of-freedom to match observed market data. Training of models requires optimizing over an SDE, which is computationally challenging. For European options, we develop a fast stochastic gradient descent (SGD) algorithm for training the neural network-SDE model. Our SGD algorithm uses two independent SDE paths to obtain an unbiased estimate of the direction of steepest descent. For American options, we optimize over the corresponding Kolmogorov partial differential equation (PDE). The neural network appears as coefficient functions in the PDE. Models are trained on large datasets (many contracts), requiring either large simulations (many Monte Carlo samples for the stock price paths) or large numbers of PDEs (a PDE must be solved for each contract). Numerical results are presented for real market data including S&P 500 index options, S&P 100 index options, and single-stock American options. The neural-network-based SDE models are compared against the Black-Scholes model, the Dupire's local volatility model, and the Heston model. Models are evaluated in terms of how accurate they are at pricing out-of-sample financial derivatives, which is a core task in derivative pricing at financial institutions. Specifically, we calibrate a neural network-SDE model to market data for a financial derivative on an asset with price S t with a payoff function g ( s ) , and we then evaluate its generalization accuracy for a financial derivative on the same asset S t but with a different payoff function f ( s ) . In addition to comparing out-of-sample pricing accuracy, we evaluate the hedging performance of the neural network-SDE model.

In this paper, deep neural networks (DNNs) are applied to features extracted from Parkinsonian speech recordings to predict their perceived quality. This procedure was also used to benchmark the electroacoustic characteristics of speech amplifiers used by people afflicted with Parkinson Disease (PD). Speech recordings were obtained from 11 PD subjects and 10 normal controls, with and without the assistance of 7 different speech amplifiers, and their quality was assessed subjectively by normal hearing listeners. Mel-frequency and Gammatone frequency cepstral coefficients (MFCCs and GFCCs respectively) and their first order derivatives were extracted as features, and given as input to the DNN. Two optimizers were used to train the neural network, namely stochastic gradient descent (SGD) and Adam optimizers. The paper also shows the effect of feature reduction in enhancing the performance of the objective predictors. Experimental results showed that a trained DNN with reduced set of GFCC features outperforms other objective metrics in terms of correlation with the subjective measures.

Survey on Deep Neural Networks in Speech and Vision Systems

Contact-map of a protein sequence dictates the global topology of structural fold. Accurate prediction of the contact-map is thus essential to protein 3D structure prediction, which is particularly useful for the protein sequences that do not have close homology templates in the Protein Data Bank. We developed a new method, ResPRE, to predict residue-level protein contacts using inverse covariance matrix (or precision matrix) of multiple sequence alignments (MSAs) through deep residual convolutional neural network training. The approach was tested on a set of 158 non-homologous proteins collected from the CASP experiments and achieved an average accuracy of 50.6% in the top-L long-range contact prediction with L being the sequence length, which is 11.7% higher than the best of other state-of-the-art approaches ranging from coevolution coupling analysis to deep neural network training. Detailed data analyses show that the major advantage of ResPRE lies at the utilization of precision matrix that helps rule out transitional noises of contact-maps compared with the previously used covariance matrix. Meanwhile, the residual network with parallel shortcut layer connections increases the learning ability of deep neural network training. It was also found that appropriate collection of MSAs can further improve the accuracy of final contact-map predictions. The standalone package and online server of ResPRE are made freely available, which should bring important impact on protein structure and function modeling studies in particular for the distant- and non-homology protein targets. https://zhanglab.ccmb.med.umich.edu/ResPRE and https://github.com/leeyang/ResPRE. Supplementary data are available at Bioinformatics online.