Simple and Automatic Distributed Machine Learning on Ray

In recent years, the pace of innovations in the fields of machine learning (ML) has accelerated, researchers in SysML have created algorithms and systems that parallelizeML training over multiple devices or computational nodes. As ML models become more structurally complex, many systems have struggled to provide allround performance on a variety of models. Particularly, ML scale-up is usually underestimated in terms of the amount of knowledge and time required to map from an appropriate distribution strategy to the model. Applying parallel training systems tocomplex models adds nontrivial development overheads in addition to model prototyping, and often results in lower-than-expected performance. This thesis identifies and addresses research challenges in both usability and performance in parallel ML techniques and system implementations. The first part of this thesis presents a simple design principle, adaptive parallelism, that applies suitable parallelization techniques to model building blocks (e.g. layers) according to their specific ML properties. Following it, we derive a series of optimizations and implementations optimizing different aspects of ML parallelization. We examine them and show that they significantly boost the efficiency or scalability of ML training on clusters 2-10x in their applicable scenarios. Generalizing this methodology, this second part of this thesis formulates the ML parallelization as an end-to-end optimization problem, and seeks to solve it automatically, for two broad paradigms of ML parallelization tasks: single-node dynamicbatching and distributed ML parallelisms. We present principled representations to express the two classes of ML parallelisms, along with composable system architectures,Cavs and AutoDist, respectively. They enable rapid compositions of parallelization strategies for unseen models, improve parallelization performance, and simplify parallel ML programming. On top of them, the third part of this thesis presents an automatic parallelization framework, AutoSync, to automatically optimize synchronization strategies indata-parallel distributed training. AutoSync achieves high performance “out-of-thebox” – it navigates the space spanned by the proposed representation, and automaticallyidentifies synchronization strategies that report 1.2 - 1.6x speedups over existing hand-optimized systems, lowering the technical barrier of distributed ML and helping make it accessible to a larger community of users. Collectively, the set of techniques and systems developed in this thesis lead to the proof of the concept and the prototype implementation of an end-to-end compiler system for large-scale ML training on distributed environments.

Aim: The majority of machine learning (ML) models in healthcare are built on retrospective data, much of which is collected without explicit patient consent for use in artificial intelligence (AI) and ML applications. The primary aim of this study was to evaluate whether clinicians and scientific researchers themselves consent to provide their own data for the training of ML models. Materials and Methods: This survey was conducted through an anonymous online survey, utilizing platforms such as Telegram, LinkedIn, and Viber. The target audience comprised specific international groups, primarily Russian, German, and English-speaking, of clinicians and scientific researchers. These participants ranged in their levels of expertise and experience, from beginners to veterans. The survey centered on a singular, pivotal question: “Do You Consent to the Use of Your Biological and Private Data for Training Machine Learning and AI Models?” Respondents had the option to choose from three responses: “Yes” and “No”. Results: The survey was conducted in January 2024. A total of 119 unique and verified individuals participated in the survey. The results revealed that only 50% of respondents (63 persons) expressed consent to provide their own data for the training of ML and AI models. Conclusion: In the development of ML and AI models, particularly open-source ones, it is crucial to ascertain whether participants are willing to provide their private data. While ML algorithms can transform the nature of data, it is important to remember that the primary owner of this data is the individual. Our findings show that in 50% of the cases, even participants from scientific research and clinical backgrounds – individuals typically accountable for ensuring data quality in AI and ML model development – do not consent to the use of their data in AI and ML settings. This highlights the need for more stringent consent processes and ethical considerations in the utilization of personal data in AI and ML research.

In the area of machine learning (ML) training data optimization through the construction of compact data, the focus of this paper is presented. The concept of compact data design, aimed at creating an optimized dataset that maximizes benefits without the need to manage a vast amount of complex data, is introduced. Improvements in the methods for optimizing ML training have been incorporated into the development of artificial intelligence (AI) systems. The introduction of understanding ML training datasets as a facet of Explainable AI (XAI), comprehensible to humans, has been made. Among the methods of XAI, the evaluation of input feature importance stands out as a way to enhance the accuracy of complex ML models. The innovative method of compact data design for optimizing ML training through dataset reduction is proposed. The performance of an ML-based malware detection system, along with its variant utilizing compact data, has been assessed, demonstrating the maintenance of 99% accuracy. By applying a 76% reduced input dataset, the speed ofMLtraining with the novel compact data design could be maximized, suggesting that anMLsystem trained in this manner could achieve statistically equivalent accuracy with only 57% of the original data sample size.

Accurate estimation of reference evapotranspiration (ETo) plays a vital role in irrigation and water resource planning. The Penman–Monteith method recommended by the Food and Agriculture Organization (FAO PM56) is widely used and considered a standard to calculate ETo. However, FAO PM56 cannot be used with limited meteorological variables, so it is compulsory to choose an alternative model for ETo estimation, which requires fewer variables. This study built ten machine learning (ML) models based on multi-function, neural network, and tree-based structure against the FAO PM56 method. For this purpose, time series temperature data on a monthly scale are only used to train ML models. The developed ML models were applied to estimate ETo at different test stations and the obtained results were compared with the FAO PM56 method to verify and validate their performance in ETo estimation for the selected stations. In addition, multiple statistical indicators, including root-mean-square error (RMSE), coefficient of determination (R2), mean absolute error (MAE), Nash–Sutcliffe efficiency (NSE), and correlation coefficient (r) were calculated to compare the performance of each ML model on ETo estimation. Among the applied ML models, the ETo tree boost (TB) ML model outperformed the other ML models in estimating ETo in diverse climatic conditions based on statistical indicators (R2, NSE, r, RMSE, and MAE). Moreover, the observed R2, NSE, and r were the highest for the TB ML model, while RMSE and MAE were found to be the lowest at the study sites compared to other applied ML models. Lastly, ETo point data yielded from the TB ML model was used in an interpolation process to create monthly and annual ETo maps. Based on the ETo maps, this study suggests mainly a focus on areas with high ETo values and proper irrigation scheduling of crops to ensure water sustainability.

Batteries, recognized as effective energy storage solutions, are considered the main facilitators of the world-wide transition towards clean and renewable energy sources. Among different types of batteries, lithium-ion (Li-ion) variants offer higher energy densities and relatively longer life spans when compared to other types. Nonetheless, a primary concern with these batteries is their lifetime. Batteries undergo various degradation mechanisms under storage and use, significantly impacting their lifespan. To this end, it is crucial to predict the degradation and lifetime of Li-ion batteries under given conditions.Researchers use three main methodologies to perform battery health diagnostics and to predict the lifetime of batteries. One approach revolves around using mechanistic, first-principle electrochemical models, also known as physics-based models. If equipped with proper thermodynamic theories, such models can show promising capabilities; however, holistic degradation prediction with these models is still challenging due to the computational complexities and the multitude of parameters that need to be fine-tuned in this approach. The other common strategy is to use empirical models to predict battery degradation. These models generally entail fewer parameters to be identified and are computationally less intensive to solve. Nonetheless, empirical models can suffer from the accuracy point-of-view as they are constrained to predict degradation trends introduced by certain degradation modes. Another common technique in battery life prediction is to utilize purely data-driven methods, such as machine learning (ML) algorithms, which also have shown promising results in the literature on rapid health predictions. However, these methods require large volumes of experimental data for training and testing ML models to ensure accuracy. In addition, data-driven methods are likely to extrapolate poorly to conditions beyond their training data and are indifferent towards the underlying degradation mechanisms. Recently, physics-informed machine learning (PI-ML) methods have garnered significant attention. They integrate physics-based or empirical models (developed based on physics) with a data-driven approach and allow one to train ML models on a smaller set of experimental data.To the best of the authors’ knowledge, the performance comparison between first-principle and empirical models when integrated within PI-ML remains unclear. Therefore, in this work, we aim to compare these two models when applied to prognostics (capacity forecasting and remaining useful life prediction) of a set of Li-ion batteries. To perform this study, we generate aging data for 40 Li-ion coin cells cycled under randomized conditions. Each cell undergoes a three-step charging stage followed by a two-step discharge stage. After obtaining the aging data, we will develop two PI-ML models, one equipped with a physics-based model and another with a set of empirical models. Both PI-ML models in this work will follow the sequential integration approach, where the training data for the final PI-ML model come from both experimental and computational data, the latter of which are obtained from the physics-based or empirical models. The parameters for the physics-based and empirical models are identified from another set of experimental data. Finally, the PI-ML models will be tested with experimental data obtained at different cycling conditions. The data flow for the sequential architecture of PI-ML is shown in the attached figure. This comparative study will help identify the performance of physics-based and empirical models when integrated into PI-ML. The main performance metric considered in this work is each model’s ability to extrapolate beyond the experimental training data set, hence aiding the final PI-ML model in generalizing to conditions not covered by its experimental training data set. Figure 1

The future 6G network will be able to connect massive and various unmanned vehicles (UxVs) into the network, bringing novel requirements toward UxV network security and data privacy. Deploying machine learning (ML)-based intrusion detection on UxVs can be one promising approach. The conventional approach cannot fulfill security and privacy requirements, as the structure can be more flexible, and there may be a lack of a central node in the 6G-based UxV network. Moreover, UxV clients can exhibit high heterogeneity with extremely class-imbalanced training samples. With the presence of federated learning (FL), UxVs can collaborate to train and update ML models while preserving data privacy in the UxV security critical scenario. However, most past works focused on centralized approaches, using cloud servers to assist intrusion detection and service deployments. This article introduces an FL framework with a decentralized design for training ML models on UxVs using intrusion detection as a study case. This scheme can be more flexible, as UxV clients can collaborate to train and synchronize models without the central server. Simulation experiments were conducted to evaluate the efficiency of the proposed approach. Experimental results and theoretical analysis confirm that the proposed method outperforms baseline models with FedAvg and local-trained client models.

Machine learning (ML) training data is often scattered across disparate collections of datasets, called <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">data silos</i> . This fragmentation poses a major challenge for data-intensive ML applications: integrating and transforming data residing in different sources demand a lot of manual work and computational resources. With data privacy constraints, data often cannot leave the premises of data silos; hence model training should proceed in a decentralized manner. In this work, we present a vision of bridging traditional data integration (DI) techniques with the requirements of modern machine learning systems. We explore the possibilities of utilizing metadata obtained from data integration processes for improving the effectiveness, efficiency, and privacy of ML models. Towards this direction, we analyze ML training and inference over data silos. Bringing data integration and machine learning together, we highlight new research opportunities from the aspects of systems, representations, factorized learning, and federated learning.

Reconstructing the emergence of the modern continental crust is crucial to understand the evolution of the crust, the onset of plate tectonics, elemental cycling, and long-term climate. However, it remains highly contentious about when and how the major subaerial continental crust emerged over time. Here, we used a machine learning (ML) model (XGBoost) to reveal the exposed history of continental crust over the last 3.8 billion years ago (Ga). First, we compiled ~10,000 modern subaerial or submarine basalts with major and trace elements to train the ML model. Then, the trained ML model (with resampling) was utilized to predict and calculate the mean proportions of subaerially erupted continental basaltic rocks since 3.8 Ga. The result suggested that the subaerial proportion only reached about 50% at ~2.5 Ga, indicating the exposure of the continental crust was far from the present-day level at the end of the Archean era. On the other hand, since ~1.8 Ga, the subaerial proportion of the continental crust exhibited a dynamic balance between ~60% and 80%, reaching the present-day level.

HADA: An automated tool for hardware dimensioning of AI applications

Federated learning enables privacy-preserving and data-efficient dimension prediction and part qualification across additive manufacturing factories

Surrogate-assisted hydraulic fracture optimization workflow with applications for shale gas reservoir development: a comparative study of machine learning models

Machine learning (ML) models based on a large data set of in-cylinder flow fields of an IC engine obtained by high-speed particle image velocimetry allow the identification of relevant flow structures underlying cycle-to-cycle variations of engine performance. To this end, deep feature learning is employed to train ML models that predict cycles of high and low in-cylinder maximum pressure. Deep convolutional autoencoders are self-supervised-trained to encode flow field features in low dimensional latent space. Without the limitations ascribable to manual feature engineering, ML models based on these learned features are able to classify high energy cycles already from the flow field during late intake and the compression stroke as early as 290 crank angle degrees before top dead center ([Formula: see text]) with a mean accuracy above chance level. The prediction accuracy from [Formula: see text] to [Formula: see text] is comparable to baseline ML approaches utilizing an extensive set of engineered features. Relevant flow structures in the compression stroke are revealed by feature analysis of ML models and are interpreted using conditional averaged flow quantities. This analysis unveils the importance of the horizontal velocity component of in-cylinder flows in predicting engine performance. Combining deep learning and conventional flow analysis techniques promises to be a powerful tool for ultimately revealing high-level flow features relevant to the prediction of cycle-to-cycle variations and further engine optimization.

Enabled by a rich ecosystem of Machine Learning (ML) libraries, programming using learned models , i.e., Software-2.0 , has gained substantial adoption. However, we do not know what challenges developers encounter when they use ML libraries. With this knowledge gap, researchers miss opportunities to contribute to new research directions, tool builders do not invest resources where automation is most needed, library designers cannot make informed decisions when releasing ML library versions, and developers fail to use common practices when using ML libraries. We present the first large-scale quantitative and qualitative empirical study to shed light on how developers in Software-2.0 use ML libraries, and how this evolution affects their code. Particularly, using static analysis we perform a longitudinal study of 3,340 top-rated open-source projects with 46,110 contributors. To further understand the challenges of ML library evolution, we survey 109 developers who introduce and evolve ML libraries. Using this rich dataset we reveal several novel findings. Among others, we found an increasing trend of using ML libraries: The ratio of new Python projects that use ML libraries increased from 2% in 2013 to 50% in 2018. We identify several usage patterns including the following: (i) 36% of the projects use multiple ML libraries to implement various stages of the ML workflows, (ii) developers update ML libraries more often than the traditional libraries , (iii) strict upgrades are the most popular for ML libraries among other update kinds, (iv) ML library updates often result in cascading library updates, and (v) ML libraries are often downgraded (22.04% of cases). We also observed unique challenges when evolving and maintaining Software-2.0 such as (i) binary incompatibility of trained ML models and (ii) benchmarking ML models. Finally, we present actionable implications of our findings for researchers, tool builders, developers, educators, library vendors, and hardware vendors.

The burgeoning field of Machine Learning (ML) has revolutionized various aspects of our lives. However, the reliance on vast amounts of data, often containing personal information, raises concerns about individual privacy. Striking a balance between effective ML model training and protecting sensitive data is crucial for responsible development and ethical implementation. This paper explores the challenges and potential solutions for preserving privacy in ML training, focusing on differential privacy (DP). The advantages of implementing DP in ML training include robust protection of individual data, enabling meaningful insights from large datasets while maintaining privacy. This is essential for ethical and responsible data usage in machine learning applications. However, DP in ML training presents challenges including scalability issues and trade-offs between utility and privacy. The paper also covers the mathematical mechanisms of Laplace and Gaussian and their noise addition, followed by a comparative analysis of their efficiency within the dataset.

Function-as-a-Service (FaaS) has raised a growing interest in how to “tame” serverless computing to enable domain-specific use cases such as data-intensive applications and machine learning (ML), to name a few. Recently, several systems have been implemented for training ML models. Certainly, these research articles are significant steps in the correct direction. However, they do not completely answer the nagging question of when serverless ML training can be more cost-effective compared to traditional “serverful” computing. To help in this endeavor, we propose MLLess, a FaaS-based ML training prototype built atop IBM Cloud Functions. To boost cost-efficiency, MLLess implements two innovative optimizations tailored to the traits of serverless computing: on one hand, a significance filter, to make indirect communication more effective, and on the other hand, a scale-in auto-tuner, to reduce cost by benefiting from the FaaS sub-second billing model (often per 100 ms). Our results certify that MLLess can be 15X faster than serverful ML systems [27] at a lower cost for sparse ML models that exhibit fast convergence such as sparse logistic regression and matrix factorization. Furthermore, our results show that MLLess can easily scale out to increasingly large fleets of serverless workers.