Search NASA⌕ Search

SEARCH · Search NASA

Results for “Learning Rate”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 19 records
All sourcesNASA NTRSDOE OSTI
DOE OSTI2020

A New Approach to Distributed Hypothesis Testing and Non-Bayesian Learning: Improved Learning Rate and Byzantine-Resilience

Here, we study a setting where a group of agents, each receiving partially informative private signals, seek to collaboratively learn the true underlying state of the world (from a finite set of hypotheses) that generates their joint observation profiles. To solve this problem, we propose a distributed learning rule that differs fundamentally from existing approaches, in that it does not employ any form of “belief-averaging”. Instead, agents update their beliefs based on a min-rule. Under standard assumptions on the observation model and the network structure, we establish that each agent learns the truth asymptotically almost surely. As our main contribution, we prove that with probability 1, each false hypothesis is ruled out by every agent exponentially fast, at a network-independent rate that is strictly larger than existing rates. We then develop a computationally-efficient variant of our learning rule that is provably resilient to agents who do not behave as expected (as represented by a Byzantine adversary model) and deliberately try to spread misinformation.

42 ENGINEERING↗
DOE OSTI2025

Open Architecture for Cost Savings in Advanced Nuclear Reactors

Recently, nuclear power plant build projects in the West have run over budget due to high capital costs and schedule overruns. Compared to other sources of energy, nuclear power plants have higher capital costs. Reactors are often different at every site, resulting in a lack of standardization. Nuclear is expected to compete with other low carbon sources of energy which have lower capital costs making it essential for nuclear to develop ways of reducing costs. Strategies such as standardization, learning rates, modularization, and schedule reduction in advanced reactors can reduce nuclear costs by about 40%. Standardization as a way of cutting capital costs has been explored even in large nuclear power plants. Standardization of certain plant components can result in lower component and installation costs and higher learning from experience. Standardization can be achieved by adopting a criterion of key performance indicators and general design principles for a specific system or component such as the balance of plant. Modularization allows the construction of certain components of SMRs in a factory, which saves time, increases productivity, and encourages higher learning rates. Production learning decreases the time and the cost related to an activity. The potential for modularized components of advanced reactors to be manufactured in factories makes it conducive to achieving higher learning rates. Developing large-capacity nuclear programs through sequential builds cultivates a higher learning rate, which in effect may reduce schedule overruns. Open architecture has been identified as a way to drive standardization among advanced reactor designs and result in cost savings. Open architecture (OA) is defined as a design enabling a diverse supply chain by defining and publishing requirements of systems or equipment in functional and/or interface terms, utilizing technical standards in widespread use. Currently, the nuclear industry’s approach is to use closed architecture, making most designs proprietary. However, collaboration between various advanced reactor vendors and suppliers utilizing the concept of open architecture can result in modular and standardized architecture of subsystems or subcomponents of a nuclear power plant. Completely standardizing nuclear power plants may be impossible, however, certain common subsystems amongst the various reactor designs could be standardized and/or access a wider supply chain and leverage existing learning from other sectors. Open architecture will save time and allocate resources to the parts of the plants that have the most unique features. A key advantage of open architecture is its ability to improve production learning across advanced reactors (AR) types in the industry, by providing and utilizing the same kind of component. Sodium fast reactor (SFR), High Temperature Gas Reactor (HTGR) and Molten Salt Reactor (MSR) are the advanced reactors considered for this project. This paper aims to determine the cost savings in advanced reactor programs due to open architecture learning rate. This work is an extension of work done on light water reactor small modular reactors; the cost methodology was utilized to investigate the impact of open architecture on advanced reactors with a particular focus on sodium fast reactors. The cost data on sodium fast reactors used in the model presented the most adequate information required for the analysis.

Advanced Nuclear Reactors↗
DOE OSTI2024

Meta-Analysis of Advanced Nuclear Reactor Cost Estimations

Supporting Data can be downloaded at: https://gain.inl.gov/content/uploads/4/2024/06/INL-RPT-24-77048-R1.xlsx Nuclear energy is a critical cornerstone of the current United States clean energy supply and may play a larger role in the future in support of a transition to a net-zero economy. The current fleet of nuclear reactors predominantly consists of large light-water reactors (LWRs), while many of the reactor designs under consideration are smaller and/or different technologies. Because these new designs have not yet been built, there is a high degree of uncertainty associated with their cost. This complicates energy-planning efforts because cost projections are not always standardized, consistent, and centralized in an easily accessible location. To help support energy planning in the US, this report provides advanced nuclear cost ranges using a transparent methodology along with other relevant information that can be used to help support decision making and energy planning. The purpose of this work was to conduct a methodical process for cost evaluation using only public information that was vetted with the end-goal to provide reference cost projections for nuclear energy. To provide a solid basis for these values, the approach and assumptions are explicitly laid out throughout the report allowing any user of the data to challenge or reconsider them. Because future US nuclear-reactor costs are still unknown due to little recent observed data, the report opted to compile a comprehensive list of bottom-up estimates and evaluate averages/trends within the data to identify reference ranges. This was deemed preferable to opining on the robustness or validity of one cost estimation versus another. To that end, the work evaluated thousands of lines of cost subaccounts from several bottom-up cost estimates. A wide variety of different reactor types captured in the data are of various sizes and technologies. Some of these reactors will be representative of advanced reactors under development while others will not. Thus, the results here are dependent on the data that are available and the accuracy of the estimates that are used. Each bottom-up estimate was reviewed to determine whether it was complete. Incomplete data sets were corrected to ensure an adequate basis of cross-comparison. The report is not without limitations and should be interpreted as an initial step to develop cost ranges for nuclear technology. Ultimately, future work can build upon the methodology with refined cost estimates to reduce uncertainty. US-based overnight capital cost (OCC) estimates were compiled from extensive data sets into ranges for both large and small reactor sizes for 2030. To project the cost declines over time, learning rates were sampled from literature sources. No SMRs were previously built; hence, learning rates based on bottom-up approaches (e.g., by quantifying the impact stemming from fabrication of different components, modular work, site construction, commissioning) were prioritized. For larger reactors, actual learning rates from deployments were used to project future costs (adjusted to account for standardization or lack thereof between designs). Other costs included are fixed and variable operations and maintenance costs. The final variables were capacity factors and ramp rates to support energy planning.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗
DOE OSTI2021

Machine learning based rate optimization under geologic uncertainty

We propose a novel approach for rate optimization during a waterflood under geologic uncertainty in reservoir properties such as permeability and porosity. The traditional approach typically involves several runs of the forward simulator. This may not scale well when the optimization is to be performed at the full field-level and over multiple geologic realizations. A machine-learning (ML) based approach which is quick and scalable for rate optimization over multiple geologic realizations is proposed instead. The training data for the model is generated by running the forward simulator with randomly assigned well rates using multiple geologic realizations. A reduced order representation of the permeability heterogeneity in each of the realizations is derived using a grid connectivity transformation (GCT). This step involves finding basis functions corresponding to the different modal frequencies of the grid connectivity represented by the grid Laplacian. The projection of the heterogeneous property field along these basis functions gives the basis coefficients that form the reduced order representation. Subsequently, for each training datapoint, streamlines are traced and the minimum time of flight (TOF) representing the tracer breakthrough time at each producer is recorded. The basis coefficients and well rates are fed to a machine learning model as input and the minimum TOF at the producers forms the output of the model. This trained model can then be used along with an optimizer for computing the optimal injection rates to maximize the injection sweep efficiency. This corresponds to minimizing the variance in the minimum TOF within each well group. Different architectures of neural network are tested using 5-fold cross validation to decide the best ML model to compute the streamline time of flight. The trained model is used to perform well rate optimization over multiple realizations of geology by using a risk tolerance penalty. The optimal well rates thus obtained are compared with two cases: a) equal well rates assigned to all injectors and producers and b) well rates obtained by optimizing over a single realization without considering the uncertainty in geology. The optimal well rates are seen to offer better oil recovery and sweep efficiency than both cases.

02 PETROLEUM↗
DOE OSTI2022

A Systematic Framework for Projecting the Future Cost of Offshore Wind Energy

Offshore wind costs are expected to decline rapidly in the short and medium term future as the industry grows and gains experience in manufacturing, installing, and operating commercial scale projects. Estimating the future costs of offshore wind energy is critical for evaluating the technology's economic performance, how it can fit into a broader clean energy economy, and how R&D investment can be allocated to advance the technology. We present a newly developed approach for forecasting these costs which focuses on an empirically-derived learning rate for capital costs and prescribed cost and performance improvements for operational costs and capacity factor. We establish baseline costs for a series of reference fixed-bottom and floating projects in 2021 and project cost trajectories to 2035, presenting both an average cost trajectory as well as describing the range of potential future costs associated with site-specific cost variations and uncertainty in the estimate of the learning rate. We also conduct sensitivity analyses showing the impact of different global deployments by 2035 and variations in the prescribed operational costs and capacity factors. The results show that fixed-bottom and floating offshore wind capital costs could decrease to around $\$$2,400/kW and $\$$3,300/kW by 2035, respectively, with ranges of $\$$2,100/kW - $\$$2,750/kW for fixed-bottom projects and $\$$2,850/kW - $\$$5,500/kW for floating projects. The levelized cost of energy of fixed-bottom and floating wind projects could decrease to $\$$53.1/MWh and $\$$63.9/MWh by 2035, with ranges of $\$$48.4/MWh - $\$$59.7/MWh for fixed-bottom projects and $\$$46.5/MWh - $\$$99.9/MWh for floating projects. By presenting the uncertainty associated with the forecast we provide a transparent description of the spectrum of potential cost trajectories for offshore wind.

17 WIND ENERGY↗
DOE OSTI2020

Locally adaptive activation functions with slope recovery for deep and physics-informed neural networks

Here we propose two approaches of locally adaptive activation functions namely, layer-wise and neuron-wise locally adaptive activation functions, which improve the performance of deep and physics-informed neural networks. The local adaptation of activation function is achieved by introducing a scalable parameter in each layer (layer-wise) and for every neuron (neuron-wise) separately, and then optimizing it using a variant of stochastic gradient descent algorithm. In order to further increase the training speed, an activation slope-based slope recovery term is added in the loss function, which further accelerates convergence, thereby reducing the training cost. On the theoretical side, we prove that in the proposed method, the gradient descent algorithms are not attracted to sub-optimal critical points or local minima under practical conditions on the initialization and learning rate, and that the gradient dynamics of the proposed method is not achievable by base methods with any (adaptive) learning rates. We further show that the adaptive activation methods accelerate the convergence by implicitly multiplying conditioning matrices to the gradient of the base method without any explicit computation of the conditioning matrix and the matrix–vector product. The different adaptive activation functions are shown to induce different implicit conditioning matrices. Furthermore, the proposed methods with the slope recovery are shown to accelerate the training process.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2020

Learning and meta-learning of stochastic advection–diffusion–reaction systems from sparse measurements

Physics-informed neural networks (PINNs) were recently proposed in [18] as an alternative way to solve partial differential equations (PDEs). A neural network (NN) represents the solution, while a PDE-induced NN is coupled to the solution NN, and all differential operators are treated using automatic differentiation. Here, we first employ the standard PINN and a stochastic version, sPINN, to solve forward and inverse problems governed by a non-linear advection–diffusion–reaction (ADR) equation, assuming we have some sparse measurements of the concentration field at random or pre-selected locations. Subsequently, we attempt to optimise the hyper-parameters of sPINN by using the Bayesian optimisation method (meta-learning) and compare the results with the empirically selected hyper-parameters of sPINN. In particular, for the first part in solving the inverse deterministic ADR, we assume that we only have a few high-fidelity measurements, whereas the rest of the data is of lower fidelity. Hence, the PINN is trained using a composite multi-fidelity network, first introduced in [12], that learns the correlations between the multi-fidelity data and predicts the unknown values of diffusivity, transport velocity and two reaction constants as well as the concentration field. For the stochastic ADR, we employ a Karhunen–Loève (KL) expansion to represent the stochastic diffusivity, and arbitrary polynomial chaos (aPC) to represent the stochastic solution. Correspondingly, we design multiple NNs to represent the mean of the solution and learn each aPC mode separately, whereas we employ a separate NN to represent the mean of diffusivity and another NN to learn all modes of the KL expansion. For the inverse problem, in addition to stochastic diffusivity and concentration fields, we also aim to obtain the (unknown) deterministic values of transport velocity and reaction constants. The available data correspond to 7spatial points for the diffusivity and 20 space–time points for the solution, both sampled 2000 times. We obtain good accuracy for the deterministic parameters of the order of 1–2% and excellent accuracy for the mean and variance of the stochastic fields, better than three digits of accuracy. In the second part, we consider the previous stochastic inverse problem, and we use Bayesian optimisation to find five hyper-parameters of sPINN, namely the width, depth and learning rate of two NNs for learning the modes. Here, we obtain much deeper and wider optimal NNs compared to the manual tuning, leading to even better accuracy, i.e., errors less than 1% for the deterministic values, and about an order of magnitude less for the stochastic fields.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2022

BUTTER - Empirical Deep Learning Dataset

The BUTTER Empirical Deep Learning Dataset represents an empirical study of the deep learning phenomena on dense fully connected networks, scanning across thirteen datasets, eight network shapes, fourteen depths, twenty-three network sizes (number of trainable parameters), four learning rates, six minibatch sizes, four levels of label noise, and fourteen levels of L1 and L2 regularization each. Multiple repetitions (typically 30, sometimes 10) of each combination of hyperparameters were preformed, and statistics including training and test loss (using a 80% / 20% shuffled train-test split) are recorded at the end of each training epoch. In total, this dataset covers 178 thousand distinct hyperparameter settings ("experiments"), 3.55 million individual training runs (an average of 20 repetitions of each experiments), and a total of 13.3 billion training epochs (three thousand epochs were covered by most runs). Accumulating this dataset consumed 5,448.4 CPU core-years, 17.8 GPU-years, and 111.2 node-years.

Array↗
DOE OSTI2026

Gravity Well Commercial Economics Assessment: Potential Revenue and Cost: Cooperative Research and Development (Final Report)

In the Gravity Well Revenue Study, we evaluate the potential revenue from energy storage using historical energy-only electricity prices, forward-looking projections of hourly electricity prices, and actual reported revenue. This analysis examines the impact of storage duration and round-trip efficiency, as well as the location of the storage, on storage revenue within the current and projected U.S. power system. We also investigated the impact of round-trip efficiency on storage revenue. We found that the relationship between storage revenue and round-trip efficiency is nonlinear. The value of improved round-trip efficiency declines as round-trip efficiency increases. In the Gravity Well Future Cost Study, we applied learning curves to predict the future cost trajectory of Gravity Wells (GrWs). Two types of analysis were implemented. The first was a bottom-up analysis that used historical learning rates for cost components, such as motors and gearboxes, and cost categories (e.g., engineering and design, etc.) to determine the learning-by-doing based single-factor learning curve. The single factor learning curve expresses the relationship between the cost of GrW and the number of units deployed (or the cumulative capacity). In the second analysis, we predicted future GrW costs via a top-down approach. This approach accounts for historical cost trends in other renewable energy and storage technologies, which have similarities with GrWs. Using a multifactor learning curve that accounts for both intrinsic (cumulative capacity) and extrinsic (the elasticity in the price of steel) factors, we estimated the future cost of GrWs.

25 ENERGY STORAGE↗
DOE OSTI2020

On Transfer Learning Of Neural Networks Using Bi-Fidelity Data For Uncertainty Propagation

Due to their high degree of expressiveness, neural networks have recently been used as surrogate models for mapping inputs of an engineering system to outputs of interest. Once trained, neural networks are computationally inexpensive to evaluate and remove the need for repeated evaluations of computationally expensive models in uncertainty quantification applications. However, given the highly parameterized construction of neural networks, especially deep neural networks, accurate training often requires large amounts of simulation data that may not be available in the case of computationally expensive systems. In this paper, to alleviate this issue for uncertainty propagation, we explore the application of transfer learning techniques using training data generated from both high- and low-fidelity models. We explore two strategies for coupling these two datasets during the training procedure, namely, the standard transfer learning and the bi-fidelity-weighted learning. In the former approach, a neural network model mapping the inputs to the outputs of interest is trained based on the low-fidelity data. The high-fidelity data are then used to adapt the parameters of the upper layer(s) of the low-fidelity network, or train a simpler neural network to map the output of the low-fidelity network to that of the high-fidelity model. In the latter approach, the entire low-fidelity network parameters are updated using data generated via a Gaussian process model trained with a small high-fidelity dataset. The parameter updates are performed via a variant of stochastic gradient descent with learning rates given by the Gaussian process model. Using three numerical examples, we illustrate the utility of these bi-fidelity transfer learning methods where we focus on accuracy improvement achieved by transfer learning over standard training approaches.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2021

Distributed Inference with Sparse and Quantized Communication

Here, we consider the problem of distributed inference where agents in a network observe a stream of private signals generated by an unknown state, and aim to uniquely identify this state from a finite set of hypotheses. We focus on scenarios where communication between agents is costly, and takes place over channels with finite bandwidth. To reduce the frequency of communication, we develop a novel event-triggered distributed learning rule that is based on the principle of diffusing low beliefs on each false hypothesis. Building on this principle, we design a trigger condition under which an agent broadcasts only those components of its belief vector that have adequate innovation, to only those neighbors that require such information. We prove that our rule guarantees convergence to the true state exponentially fast almost surely despite sparse communication, and that it has the potential to significantly reduce information flow from uninformative agents to informative agents. Next, to deal with finite-precision communication channels, we propose a distributed learning rule that leverages the idea of adaptive quantization. We show that by sequentially refining the range of the quantizers, every agent can learn the truth exponentially fast almost surely, while using just 1 bit to encode its belief on each hypothesis. For both our proposed algorithms, we rigorously characterize the trade-offs between communication-efficiency and the learning rate.

42 ENGINEERING↗
DOE OSTI2024

Towards provably efficient quantum algorithms for large-scale machine-learning models

Large machine learning models are revolutionary technologies of artificial intelligence whose bottlenecks include huge computational expenses, power, and time used both in the pre-training and fine-tuning process. In this work, we show that fault-tolerant quantum computing could possibly provide provably efficient resolutions for generic (stochastic) gradient descent algorithms, scaling as $\mathcal{O}$(T 2 x polylog($n$)), where n is the size of the models and T is the number of iterations in the training, as long as the models are both sufficiently dissipative and sparse, with small learning rates. Based on earlier efficient quantum algorithms for dissipative differential equations, we find and prove that similar algorithms work for (stochastic) gradient descent, the primary algorithm for machine learning. In practice, we benchmark instances of large machine learning models from 7 million to 103 million parameters. We find that, in the context of sparse training, a quantum enhancement is possible at the early stage of learning after model pruning, motivating a sparse parameter download and re-upload scheme. Our work shows solidly that fault-tolerant quantum algorithms could potentially contribute to most state-of-the-art, large-scale machine-learning problems.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Surrogate Model Based Optimization for Finding Robust Deep Learning Model Architectures

Deep Learning (DL) models are increasingly used throughout the sciences. However, their performance and usefulness depend greatly on their architecture which is defined by hyperparameters such as the number of nodes, layers, the learning rate, etc. Tuning these hyperparameters is time-consuming because evaluating their performance requires a lengthy training step. Stochastic optimizers used in training lead to performance variability and potentially prediction reliability issues. In this talk, we will describe an automated optimization method based on surrogate models and active learning strategies for tuning DL model architectures. We take into account the prediction variability with the goal to identify architectures that make reliable and robust predictions. We demonstrate our developments on an application arising in particle physics.

deep learning↗
DOE OSTI2020

Knowledge Spillovers and Cost Reductions in Solar Soft Costs

Despite the commonly acknowledged importance of knowledge spillovers in reducing solar soft costs, we are only beginning to answer a fundamental question: who learns what (knowledge acquisition), from whom (knowledge production), and how (spillover mechanisms)? Until recently, this important topic has been largely unexplored in the case of solar soft costs. Thus, this project set out to identify how knowledge spillovers affect soft costs in the U.S. photovoltaic (PV) installation industry, specifically how important spillovers are, what types of knowledge are most likely to spillover, and how networks of actors affect spillovers. Our findings offer insights for designing solutions that address problems associated with knowledge spillovers and that leverage spillovers to reduce solar soft costs. Recognizing the ambiguity in the definition of soft costs, i.e., “non-hardware costs,” and variability in soft cost categories, we developed the Solar Soft Cost Ontology (SSCO) to systematically identify key concepts related to soft costs, network actors, learning processes, and the relationships between them. This ontology served as a foundational organizational structure for the methodology of the remaining tasks: case studies, surveys, pricing analysis, patent analysis, network analysis, and project integration across tasks. While there is substantial learning among installers that is reducing the soft costs for PV installations, most of that learning is retained by firms rather than spread across the industry. The positive relationship between experience accumulation and cost reductions is typically explained as learning by doing (LBD), but we find that LBD effects are mediated by other learning mechanisms, including learning by searching and learning by interacting. Knowledge spillovers have significant potential to reduce solar PV soft costs, but successful knowledge spillover pathways are complex and non-trivial. There are a wide variety of ways to construct an installation business, thus categories of firms that can effectively cross-learn directly are small and what knowledge is relevant to whom is challenging and costly for firms to assess. This fragmentation limits the critical mass needed for spillover related soft cost reductions. Knowledge does not flow directly between installers. Indirect knowledge transfer pathways are critical: distributors, software providers, collaboratives, and hiring. Furthermore, diverse, more integrated knowledge networks tend to promote successful learning by organizations and across the system as a whole. Accordingly, we find the need to supporting the whole ecosystem using an integrated policy and programmatic approach to support installers, distributors, complementary sector, and facilitators. Overall, a deliberate policy-mix design is needed to reduce the solar PV deployment barrier in terms of installation cost reductions, because deployment policies could potentially interact with policies that facilitate network-building and technological innovation. A combination of deployment policies, innovation-support policies, and network-facilitating policies could potentially lead to a more desired market outcome through achieving higher joint learning rates from firms’ cumulative experiences developed in a more integrated production and deployment ecosystem.

14 SOLAR ENERGY↗
DOE OSTI2024

Unraveling the impact of initial choices and in-loop interventions on learning dynamics in autonomous scanning probe microscopy

The current focus in Autonomous Experimentation (AE) is on developing robust workflows to conduct the AE effectively. This entails the need for well-defined approaches to guide the AE process, including strategies for hyperparameter tuning and high-level human interventions within the workflow loop. This paper presents a comprehensive analysis of the influence of initial experimental conditions and in-loop interventions on the learning dynamics of Deep Kernel Learning (DKL) within the realm of AE in scanning probe microscopy. We explore the concept of the “seed effect,” where the initial experiment setup has a substantial impact on the subsequent learning trajectory. Additionally, we introduce an approach of the seed point interventions in AE allowing the operator to influence the exploration process. Using a dataset from Piezoresponse Force Microscopy on PbTiO 3 thin films, we illustrate the impact of the “seed effect” and in-loop seed interventions on the effectiveness of DKL in predicting material properties. The study highlights the importance of initial choices and adaptive interventions in optimizing learning rates and enhancing the efficiency of automated material characterization. This work offers valuable insights into designing more robust and effective AE workflows in microscopy with potential applications across various characterization techniques.

47 OTHER INSTRUMENTATION↗
DOE OSTI2023

Improving Deep Neural Networks’ Training for Image Classification With Nonlinear Conjugate Gradient-Style Adaptive Momentum

Momentum is crucial in stochastic gradient-based optimization algorithms for accelerating or improving training deep neural networks (DNNs). In deep learning practice, the momentum is usually weighted by a well-calibrated constant. However, tuning the hyperparameter for momentum can be a significant computational burden. In this article, we propose a novel adaptive momentum for improving DNNs training; this adaptive momentum, with no momentum-related hyperparame- ter required, is motivated by the nonlinear conjugate gradient (NCG) method. Stochastic gradient descent (SGD) with this new adaptive momentum eliminates the need for the momentum hyperparameter calibration, allows using a significantly larger learning rate, accelerates DNN training, and improves the final accuracy and robustness of the trained DNNs. For example, SGD with this adaptive momentum reduces classification errors for training ResNet110 for CIFAR10 and CIFAR100 from 5.25% to 4.64% and 23.75% to 20.03%, respectively. Furthermore, SGD, with the new adaptive momentum, also benefits adversarial training and, hence, improves the adversarial robustness of the trained DNNs.

97 MATHEMATICS AND COMPUTING↗