Stable Anderson Acceleration for Deep Learning

Anderson acceleration (AA) is an extrapolation technique designed to speed-up fixed-point iterations like those arising from the iterative training of DL models. Training DL models requires large datasets processed in randomly sampled batches that tend to introduce in the fixed-point iteration stochastic oscillations of amplitude roughly inversely proportional to the size of the batch. These oscillations reduce and occasionally eliminate the positive effect of AA. To restore AA's advantage, we combine it with an adaptive moving average procedure that smoothes the oscillations and results in a more regular sequence of gradient descent updates. By monitoring the relative standard deviation between consecutive iterations, we also introduce a criterion to automatically assess whether the moving average is needed. We applied the method to the following DL instantiations: (i) multi-layer perceptrons (MLPs) trained on the open-source graduate admissions dataset for regression, (ii) physics informed neural networks (PINNs) trained on source data to solve 2d and 100d Burgers' partial differential equations (PDEs), and (iii) ResNet50 trained on the open-source ImageNet1k dataset for image classification. Numerical results obtained using up to 1,536 NVIDIA V100 GPUs on the OLCF supercomputer Summit showed the stabilizing effect of the moving average on AA for all the problems above.

Neural network based order parameter for phase transitions and its applications in high-entropy alloys

Phase transition is one of the most important phenomena in nature and plays a central role in materials design. All phase transitions are characterized by suitable order parameters, including the order-disorder phase transition. However, finding a representative order parameter for complex systems is nontrivial, such as for high-entropy alloys. Given variational autoencoder's (VAE) strength of reducing high dimensional data into few principal components, here we coin a new concept of "VAE order parameter". We propose that the Manhattan distance in the VAE latent space can serve as a generic order parameter for order-disorder phase transitions. The physical properties of the order parameter are quantitatively interpreted and demonstrated by multiple refractory high-entropy alloys. Assisted by it, a generally applicable alloy design concept is proposed by mimicking the nature mixing of elements. Our physically interpretable "VAE order parameter" lays the foundation for the understanding of and alloy design by chemical ordering.

Scalable Balanced Training of Conditional Generative Adversarial Neural Networks on Image Data

We propose a distributed approach to train deep convolutional generative adversarial neural network (DC-CGANs) models. Our method reduces the imbalance between generator and discriminator by partitioning the training data according to data labels, and enhances scalability by performing a parallel training where multiple generators are concurrently trained, each one of them focusing on a single data label. Performance is assessed in terms of inception score and image quality on MNIST, CIFAR10, CIFAR100, and ImageNet1k datasets, showing a significant improvement in comparison to state-of-the-art techniques to training DC-CGANs. Weak scaling is attained on all the four datasets using up to 1,000 processes and 2,000 NVIDIA V100 GPUs on the OLCF supercomputer Summit.

Data optimization for large batch distributed training of deep neural networks

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.

Distributed Training and Optimization Of Neural Networks

Deep learning models are yielding increasingly better performances thanks to multiple factors. To be successful, model may have large number of parameters or complex architectures and be trained on large dataset. This leads to large requirements on computing resource and turn around time, even more so when hyper-parameter optimization is done (e.g search over model architectures). While this is a challenge that goes beyond particle physics, we review the various ways to do the necessary computations in parallel, and put it in the context of high energy physics.

Integrating Deep Learning in Domain Sciences at Exascale

This paper presents some of the current challenges in designing deep learning artificial intelligence (AI) and integrating it with traditional high-performance computing (HPC) simulations. We evaluate existing packages for their ability to run deep learning models and applications on large-scale HPC systems efficiently, identify challenges, and propose new asynchronous parallelization and optimization techniques for current large-scale heterogeneous systems and upcoming exascale systems. These developments, along with existing HPC AI software capabilities, have been integrated into MagmaDNN, an open-source HPC deep learning framework. Many deep learning frameworks are targeted at data scientists and fall short in providing quality integration into existing HPC workflows. This paper discusses the necessities of an HPC deep learning framework and how those needs can be provided (e.g., as in MagmaDNN) through a deep integration with existing HPC libraries, such as MAGMA and its modular memory management, MPI, CuBLAS, CuDNN, MKL, and HIP. Advancements are also illustrated through the use of algorithmic enhancements in reduced- and mixed-precision, as well as asynchronous optimization methods. Finally, we present illustrations and potential solutions for enhancing traditional compute- and data-intensive applications at ORNL and UTK with AI. The approaches and future challenges are illustrated in materials science, imaging, and climate applications.

Exascale Deep Learning for Scientific Inverse Problems

We introduce novel communication strategies in synchronous distributed Deep Learning consisting of decentralized gradient reduction orchestration and computational graph-aware grouping of gradient tensors. These new techniques produce an optimal overlap between computation and communication and result in near-linear scaling (0.93) of distributed training up to 27,600 NVIDIA V100 GPUs on the Summit Supercomputer. We demonstrate our gradient reduction techniques in the context of training a Fully Convolutional Neural Network to approximate the solution of a longstanding scientific inverse problem in materials imaging. The efficient distributed training on a dataset size of 0.5 PB, produces a model capable of an atomically-accurate reconstruction of materials, and in the process reaching a peak performance of 2.15(4) EFLOPS$_{16}$.

A greedy constructive algorithm for the optimization of neural network architectures

In this work we propose a new method to optimize the architecture of an artificial neural network. The algorithm proposed, called Greedy Search for Neural Network Architecture, aims to minimize the complexity of the architecture search and the complexity of the final model selected without compromising the predictive performance. The reduction of the computational cost makes this approach appealing for two reasons. Firstly, there is a need from domain scientists to easily interpret predictions returned by a deep learning model and this tends to be cumbersome when neural networks have complex structures. Secondly, the use of neural networks is challenging in situations with compute/memory limitations. Promising numerical results show that our method is competitive against other hyperparameter optimization algorithms for attainable performance and computational cost. We also generalize the definition of adjusted score from linear regression models to neural networks. Numerical experiments are presented to show that the adjusted score can boost the greedy search to favor smaller architectures over larger ones without compromising the predictive performance.

Robust data-driven approach for predicting the configurational energy of high entropy alloys

High entropy alloys (HEAs) have been increasingly attractive as promising next-generation materials due to their various excellent properties. It's necessary to essentially characterize the degree of chemical ordering and identify order-disorder transitions through efficient simulation and modeling of thermodynamics. In this study, a robust data-driven framework based on Bayesian approaches is proposed and demonstrated on the accurate and efficient prediction of configurational energy of high entropy alloys. The proposed effective pair interaction (EPI) model with ensemble sampling is used to map the configuration and its corresponding energy. Given limited data calculated by first-principles calculations, Bayesian regularized regression not only offers an accurate and stable prediction but also effectively quantifies the uncertainties associated with EPI parameters. Compared with the arbitrary determination of model complexity, we further conduct a physical feature selection to identify the truncation of coordination shells in EPI model using Bayesian information criterion. The results achieve efficient and robust performance in predicting the configurational energy, particularly given small data. The developed methodology is applied to study a series of refractory HEAs, i.e. NbMoTaW, NbMoTaWV and NbMoTaWTi where it is demonstrated how dataset size affects the confidence we can place in statistical estimates of configurational energy when data are sparse.