Search NASA⌕ Search

SEARCH · Search NASA

Results for “differentiable machine learning”

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 307 records · Page 17
All sourcesNASA NTRSDOE OSTI
DOE OSTI2023

$\text{GPLaSDI}$: Gaussian Process-based interpretable Latent Space Dynamics Identification through deep autoencoder

Numerically solving partial differential equations (PDEs) can be challenging and computationally expensive. This has led to the development of reduced-order models (ROMs) that are accurate but faster than full order models (FOMs). Recently, machine learning advances have enabled the creation of non-linear projection methods, such as Latent Space Dynamics Identification (LaSDI). LaSDI maps full-order PDE solutions to a latent space using autoencoders and learns the system of ODEs governing the latent space dynamics. By interpolating and solving the ODE system in the reduced latent space, fast and accurate ROM predictions can be made by feeding the predicted latent space dynamics into the decoder. In this paper, we introduce GPLaSDI, a novel LaSDI-based framework that relies on Gaussian process (GP) for latent space ODE interpolations. Using GPs offers two significant advantages. First, it enables the quantification of uncertainty over the ROM predictions. Second, leveraging this prediction uncertainty allows for efficient adaptive training through a greedy selection of additional training data points. This approach does not require prior knowledge of the underlying PDEs. Consequently, GPLaSDI is inherently non-intrusive and can be applied to problems without a known PDE or its residual. Here we demonstrate the effectiveness of our approach on the Burgers equation, Vlasov equation for plasma physics, and a rising thermal bubble problem. Our proposed method achieves between 200 and 100,000 times speed-up, with up to 7% relative error.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2021

Privacy-Preserving Knowledge Transfer with Bootstrap Aggregation of Teacher Ensembles

There is a need to transfer knowledge among institutions and organizations to save effort in annotation and labeling or in enhancing task performance. However, knowledge transfer is difficult because of restrictions that are in place to ensure data security and privacy. Institutions are not allowed to exchange data or perform any activity that may expose personal information. With the leverage of a differential privacy algorithm in a high-performance computing environment, we propose a new training protocol, Bootstrap Aggregation of Teacher Ensembles (BATE), which is applicable to various types of machine learning models. The BATE algorithm is based on and provides enhancements to the PATE algorithm, maintaining competitive task performance scores on complex datasets with underrepresented class labels.We conducted a proof-of-the-concept study of the information extraction from cancer pathology report data from four cancer registries and performed comparisons between four scenarios: no collaboration, no privacy-preserving collaboration, the PATE algorithm, and the proposed BATE algorithm. The results showed that the BATE algorithm maintained competitive macro-averaged F1 scores, demonstrating that the suggested algorithm is an effective yet privacy-preserving method for machine learning and deep learning solutions.

Yoon, Hong-Jun↗
DOE OSTI2018

OPTIMIZATION OF FAST FISSION GAS RELEASE MODEL PARAMETERS USING MACHINE LEARNING ACCELERATED EVOLUTIONARY ALGORITHMS

Fission gas release models used in FRAPCON and FAST are based on equations that describe the physical transport phenomena that control release from the fuel pellet to the rod void volume of fission gas produced during irradiation. Although these models are physically based, many of the parameters that control the release such as the impact of burnup on diffusion and bubble formation and saturation are critical to the accurate prediction of fission gas release, but cannot be measured in any quantitative way. To overcome this problem, these parameters have been empirically derived to provide a best fit to the available data that includes rod puncture data and electron probe microanalysis (EMPA) and X-ray fluorescence (XRF) data of radial location of residual gas within the pellet. In general, a heuristic approach for deriving these parameters via practical implementation of expert analyses on just a few dependent variables has been shown to perform quite well in most situations. In this paper, we present an approach for applying machine learning to expedite model development. This approach is based on developing a deep artificial neural network which describes FRAPCON’s fission gas release models, and optimizes parameters by a differential evaluation algorithm. This approach allows us to quickly and accurately tune physical models based on expert judgment, and works as a human-in-the-loop approach, to assist the modeler in identifying and addressing regions of high uncertainty in a multi-parameter space. Results of the updated fission gas release model will be shown for all the FAST assessment data.

Johns, Jesse M.↗
DOE OSTI2024

Efficient Anomaly Detection Driven By Different Machine Learning Architectures And Models

The rapid growth and ubiquitous adoption of the internet and cyber-physical systems (CPS) have fundamentally transformed modern communication, work, and human-system interactions. While networks now form the backbone of critical digital ecosystems, enabling seamless data transmission across diverse, interconnected systems, this increased connectivity also expands the attack surface, making real-time detection of network intrusions and anomalies a pressing challenge. Detecting unusual activities within network infrastructure requires advanced data traffic analysis to differentiate between legitimate and malicious interactions. Traditional approaches to network anomaly detectionâ??such as rule-based and signature-based systemsâ??often depend on predefined patterns to identify known anomalies, limiting their effectiveness against emerging, stealthy, or previously unseen threats. These conventional methods suffer from high false alarm rates and fail to adapt to the ever-evolving nature of network traffic, particularly in large-scale, decentralized environments where data volume, velocity, and variety are constantly increasing. This dissertation presents artificial intelligence (AI)-driven approaches to anomaly detection that leverage graphics processing unit (GPU)-enabled high-performance computing (HPC) platforms for processing massive network traffic data and monitoring the components of cyber-physical systems (CPS) for potentially hazardous conditions. The research advances several key contributions: (1) Designing efficient machine learning techniques for CPS condition monitoring and anomaly detection; (2) enabling federated learning (FL) frameworks that enable distributed detection while preserving data privacy and system resilience; (3) exploring graph-based methodologies combining graph neural networks (GNN) and graph machine learning (ML) approaches for the Internet of Things (IoT) and automotive network security, and (4) performing distributed edge computing optimizations that integrate FL with scalable technologies for reduced communication overhead. Through extensive experiments, these methodologies demonstrate that complex anomaly detection and condition monitoring tasks can be achieved while balancing computational efficiency and detection accuracy through fine-grained network information processing. The frameworks developed in this research establish a robust foundation for network anomaly detection, providing scalable, adaptive, and privacy-preserving solutions for safeguarding CPS and IoT networks in an increasingly interconnected digital landscape. The practical implications of these research findings are significant, as they can inform the development of next-generation network security systems and contribute to the protection of critical infrastructure against sophisticated cyber attacks.

Marfo, William↗
DOE OSTI2022

Learning generative neural networks with physics knowledge

Deep generative neural networks have enabled modeling complex distributions, but incorporating physics knowledge into the neural networks is still challenging and is at the core of current physics-based machine learning research. To this end, we propose a physics generative neural network (PhysGNN), a new class of generative neural networks for learning unknown distributions in a physical system described by partial differential equations (PDE). PhysGNN couples PDE systems with generative neural networks. It is a fully differentiable model that allows back-propagation of gradients through both numerical PDE solvers and generative neural networks, and is trained by minimizing the discrete Wasserstein distance between generated and observed probability distributions of the PDE outputs using the stochastic gradient descent method. Moreover, PhysGNN does not require adversarial training like standard generative neural networks, which offers better stability than adversarial training. We show that PhysGNN can learn complex distributions in stochastic inverse problems, where conventional methods such as maximum likelihood estimation and momentum matching methods may be inapplicable when little knowledge is known about the form of unknown distributions or the physical model is too complex. Furthermore, our method allows physics-based generative neural network training for learning complex distributions in the context of differential equations.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

Leveraging differentiable programming in the inverse problem of neutron stars

Neutron stars (NSs) probe the high-density regime of the nuclear equation of state (EOS). However, inferring the EOS from observations of NSs is a computationally challenging task. Here, in this work, we efficiently solve this inverse problem by leveraging differential programming in two ways. First, we enable full Bayesian inference in under one hour of wall time on a GPU by using gradient-based samplers, without requiring pretrained machine learning emulators. Moreover, we demonstrate efficient scaling to high-dimensional parameter spaces. Second, we introduce a novel gradient-based optimization scheme that recovers the EOS of a given NS mass-radius curve. We demonstrate how our framework can reveal consistencies or tensions between nuclear physics and astrophysics. First, we show how the breakdown density of a metamodel description of the EOS can be determined from NS observations. Second, we demonstrate how degeneracies in EOS modeling using nuclear empirical parameters can influence the inverse problem during gradient-based optimization. Looking ahead, our approach opens up new theoretical studies of the relation between NS properties and the EOS, while effectively tackling the data analysis challenges brought by future detectors.

73 NUCLEAR PHYSICS AND RADIATION PHYSICS↗
DOE OSTI2026

Toward an event-level analysis of hadron structure using differential programming

Reconstructing the internal properties of hadrons in terms of fundamental quark and gluon de- grees of freedom is a central goal in nuclear and particle physics. This effort lies at the core of major experimental programs, such as the Jefferson Lab 12 GeV program and the upcoming Electron-Ion Collider. A primary challenge is the inherent inverse problem: converting large-scale observational data from collision events into the fundamental QCD-defined densities that characterize the micro- scopic structure of hadronic systems. Recent advances in AI and machine learning have opened new avenues for addressing this challenge using deep learning techniques. A particularly promising direction is the integration of complex theoretical calculations and experimental simulations into a unified framework capable of reconstructing these densities directly from event-level information. In this document, we introduce a key algorithm called LOITS, which enables differentiable program- ming within such a framework, facilitating the use of AI/ML techniques to solve the inverse problem of QCF reconstruction at the event level.

Braga, Kevin [College of William and Mary, William↗
DOE OSTI2022

Discovering equations that govern experimental materials stability under environmental stress using scientific machine learning

Abstract While machine learning (ML) in experimental research has demonstrated impressive predictive capabilities, extracting fungible knowledge representations from experimental data remains an elusive task. In this manuscript, we use ML to infer the underlying differential equation (DE) from experimental data of degrading organic-inorganic methylammonium lead iodide (MAPI) perovskite thin films under environmental stressors (elevated temperature, humidity, and light). Using a sparse regression algorithm, we find that the underlying DE governing MAPI degradation across a broad temperature range of 35 to 85 °C is described minimally by a second-order polynomial. This DE corresponds to the Verhulst logistic function, which describes reaction kinetics analogous to self-propagating reactions. We examine the robustness of our conclusions to experimental variance and Gaussian noise and describe the experimental limits within which this methodology can be applied. Our study highlights the promise and challenges associated with ML-aided scientific discovery by demonstrating its application in experimental chemical and materials systems.

36 MATERIALS SCIENCE↗
DOE OSTI2020

Non-autoregressive time-series methods for stable parametric reduced-order models

Advection-dominated dynamical systems, characterized by partial differential equations, are found in applications ranging from weather forecasting to engineering design where accuracy and robustness are crucial. There has been significant interest in the use of techniques borrowed from machine learning to reduce the computational expense and/or improve the accuracy of predictions for these systems. These rely on the identification of a basis that reduces the dimensionality of the problem and the subsequent use of time series and sequential learning methods to forecast the evolution of the reduced state. Often, however, machine-learned predictions after reduced-basis projection are plagued by issues of stability stemming from incomplete capture of multiscale processes as well as due to error growth for long forecast durations. To address these issues, we have developed a non-autoregressive time series approach for predicting linear reduced-basis time histories of forward models. In particular, we demonstrate that non-autoregressive counterparts of sequential learning methods such as long short-term memory (LSTM) considerably improve the stability of machine-learned reduced-order models. Further, we evaluate our approach on the inviscid shallow water equations and show that a non-autoregressive variant of the standard LSTM approach that is bidirectional in the principal component directions obtains the best accuracy for recreating the nonlinear dynamics of partial observations. Moreover-and critical for many applications of these surrogates-inference times are reduced by three orders of magnitude using our approach, compared with both the equation-based Galerkin projection method and the standard LSTM approach.

97 MATHEMATICS AND COMPUTING↗
NASA NTRS1991

An introduction to autonomous control systems

The functions, characteristics, and benefits of autonomous control are outlined. An autonomous control functional architecture for future space vehicles that incorporates the concepts and characteristics described is presented. The controller is hierarchical, with an execution level (the lowest level), coordination level (middle level), and management and organization level (highest level). The general characteristics of the overall architecture, including those of the three levels, are explained, and an example to illustrate their functions is given. Mathematical models for autonomous systems, including 'logical' discrete event system models, are discussed. An approach to the quantitative, systematic modeling, analysis, and design of autonomous controllers is also discussed. It is a hybrid approach since it uses conventional analysis techniques based on difference and differential equations and new techniques for the analysis of the systems described with a symbolic formalism such as finite automata. Some recent results from the areas of planning and expert systems, machine learning, artificial neural networks, and the area restructurable controls are briefly outlined.

Antsaklis, Panos J.↗
DOE OSTI2021

Process Anomaly Detection for Sparsely Labeled Events in Nuclear Power Plants

An essential aspect of online monitoring, subtle anomaly detection increases the detection lead time for equipment failure and enables a nuclear power plant (NPP) to mitigate unexpected partial or full outages, resulting in significant cost saving to the plant. Once an anomaly is detected by plant staff, its cause and severity are investigated. Because the vast majority of anomalies require some level of investigation, including some that require time-consuming examination, before they are passed over to the engineering organization for further analysis, plants are often equipped with tools to assist the staff in performing anomaly detection. Those tools operate as a black box and are often based on statistical methods that establish sensor correlations using preconfigured mathematical models and flag correlation deviations as anomalies. Due to the number of anomalies detected at a given NPP on a daily basis, a significant number of flagged anomalies usually await examination for days or weeks. A primary cause of this backlog is that the methods used by the tools generate many false positives. Though this is usually attributed to oversensitive model settings due to very narrow normal operation bands, it can also be associated with the model development being inadequate for the process being monitored, or with missing model inputs that could have explained misclassified positives. The performance of anomaly detection tools impacts their plant acceptance and utilization, especially when the effort to address false positives generated by the tool depletes the value or cost saved by using that tool. Thus, means to advance anomaly detection performance have been investigated by the Department of Energy’s Light Water Reactor Sustainability program. Previous and ongoing efforts have targeted unsupervised machine-learning (ML) methods, which do not require the labeling of any data fed into the ML model. By contrast, in supervised anomaly detection methods, every data point is labeled as either a normal or abnormal process condition, and the model is trained to replicate the classification process. Supervised methods usually outperform unsupervised methods, due to the added value in differentiating normal from anomalous states of the monitored process. An NPP’s corrective action program requires it to track and document, via a dedicated report, the resolution of any issues that occur within the plant. Once created, each report is reviewed by a plant screening committee, and several classifications and decisions are made. Recently, a collaborating NPP developed an artificial intelligence and ML-based classifier to categorize a condition report (CR) into classes that can serve to label the data as normal or anomalous. Applying CRs as labels represents a semi-supervised use case. Semi-supervised ML assumes that labels exist for some data points (i.e., labeled anomalies, in this case) but not for the rest. In this effort, semi-supervised ML methods were used to fuse data from CRs with anomaly detection methods in order to test the hypothesis that partially labeled anomalies would improve the accuracy of the anomaly detection methods. Specifically, two methods were used. The first is the deep Semi-supervised Anomaly Detection (deep SAD) method, which can handle labels ranging from fully unsupervised to fully supervised cases. The second is a newly designed ML method developed specifically for this effort and referred to as the high-order feature (HOF)-based method. To evaluate these two methods in controlled environments, synthetic data generators were developed and used. The first datasets used a spring-mass-damper (SMD) system simulator commonly found in mechanical engineering references. This was used to create two use cases: a one- and a three-mass system. Anomalies were introduced by changing the spring and damper coefficients while the system was actuated by random forces. The second datasets used the commercial Dymola-Modelica software to build a simplified nuclear reactor model. Anomalies were added in the form of corrupted sensor readings and/or control commands. The deep SAD method was tested using the SMD system, while the HOF method was tested using both datasets. Application of the deep SAD semi-supervised ML method demonstrated that labels can generate increased confidence in detecting true anomalies. This helped increase the number of true positives and decrease the number of false negatives—something that would aid in addressing the backlog of possible anomalies. Application of the HOF method demonstrated that labels can aid in down selecting from a candidate set of features to a more optimal subset in order to better differentiate between normal and anomalous conditions.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗
DOE OSTI2025

Accelerating the Structure Exploration of Diverse Bi–Pt Nanoclusters via Physics‐Informed Machine Learning Potential and Particle Swarm Optimization

Bimetallic Bi–Pt nanoclusters exhibit diverse structural motifs, including core-shell, Janus, and mixed alloy configurations, due to the unique bonding characteristics between Bi and Pt atoms. Using density functional theory refinements from ChIMES physically machine-learned potential and CALYPSO particle swarm optimization global searches, 34 Bi20-Pt20 nanoclusters are systematically classified. The results reveal that Bi atoms predominantly occupy surface sites, driven by charge transfer effects. Cohesive energy trends alone prove insufficient for structure differentiation, necessitating a data-driven approach employing principal component analysis and K-means clustering. Furthermore, vibrational, electronic, and infrared spectral analyses provide additional insights into structure-property relationships. The findings offer an original framework for the automated classification and analysis of bimetallic nanoclusters, enhancing the understanding of their stability and functional properties.

bimetallic nanoparticles↗
DOE OSTI2022

Creating ground truth for nanocrystal morphology: a fully automated pipeline for unbiased transmission electron microscopy analysis

Control over colloidal nanocrystal morphology (size, size distribution, and shape) is important for tailoring the functionality of individual nanocrystals and their ensemble behavior. Despite this, traditional methods to quantify nanocrystal morphology are laborious. New developments in automated morphology classification will accelerate these analyses but the assessment of machine learning models is limited by human accuracy for ground truth, causing even unsupervised machine learning models to have inherent bias. Herein, we introduce synthetic image rendering to solve the ground truth problem of nanocrystal morphology classification. By simulating 2D images of nanocrystal shapes via a function of high-dimensional parameter space, we trained a convolutional neural network to link unique morphologies to their simulated parameters, defining nanocrystal morphology quantitatively rather than qualitatively. An automated pipeline then processes, quantitatively defines, and classifies nanocrystal morphology from experimental transmission electron microscopy (TEM) images. Using improved computer vision techniques, 42,650 nanocrystals were identified, assessed, and labeled with quantitative parameters, offering a 600-fold improvement in efficiency over best-practice manual measurements. Further, a classification algorithm was trained with a prediction accuracy of 99.5%, which can successfully analyze a range of concave, convex, and irregular nanocrystal shapes. The resulting pipeline was applied to differentiating two syntheses of nominally cuboidal CsPbBr 3 nanocrystals and uniquely classifying binary nickel sulfide nanocrystal phase based on morphology. This pipeline provides a simple, efficient, and unbiased method to quantify nanocrystal morphology and represents a practical route to construct large datasets with an absolute ground truth for training unbiased morphology-based machine learning algorithms.

77 NANOSCIENCE AND NANOTECHNOLOGY↗
DOE OSTI2022

Context-Aware Learning for Inverse Design in Photovoltaics

This document describes progress in the ARPA-E DIFFERENTIATE project titled “Context-Aware Learning for Inverse Design in Photovoltaics” during the period of May 2019 to May 2022. This project is being performed at Iowa State University, New York University, Stanford University, and National Renewable Energy Laboratory. The project aims to develop a new machine learning (ML) framework to significantly accelerate the design of organic microstructures for improved organic photovoltaic performance. In this project, we had developed an inverse design framework using Deep Learning called InvNets for generating microstructures with desired physics-driven properties. As a preliminary product, in Milestone 3, we demonstrated how InvNets show 20% improvement in the performance of the microstructures and over 100X speedup in the performance compared to traditional processes for physics-driven inverse design. Later, in Milestone 6, we demonstrated that InvNets work for more complex physics properties, specifically, generating microstructures for organic photovoltaic cells with desired current-voltage characteristics. Further, in Milestone 4, we explored the idea of using physics-aware surrogates for obtaining solutions of partial differential equations(PDE) called as DiffNets(now called as NeuFENets to avoid ambiguity of names). The connection between both frameworks is that DiffNet surrogates form the physics-aware surrogate in the InvNet framework. Finally in Milestone 8, we extend our framework for other physics domains. Specifically, we explore building geometry-aware NeuFENets by developing physics surrogates that exploit ideas from traditional immersed boundary finite element methods. With these updates, we are able to achieve all the Milestones.

36 MATERIALS SCIENCE↗
DOE OSTI2026

3D reconstruction and neural rendering for adversarial machine learning

While evasion attacks on computer vision systems have been widely studied, creating attacks that remain effective under significant changes in viewpoint continues to be challenging. Traditional approaches often rely on affine transformations of images, but these approaches degrade at larger perspective shifts and often produce unrealistic or ineffective perturbations. Recent methods use differentiable renderers to improve viewpoint robustness, but they typically depend on manually constructed 3D models. We introduce a semi-automated pipeline that generates physically printable and perspective-invariant adversarial patches using only a small set of 2D images. Our method integrates 3D reconstruction, neural rendering, adversarial patch optimization, and an object detection victim model into a unified workflow. We use 2D Gaussian Splatting for high fidelity mesh reconstruction and FlexPara for surface parameterization that produces texture maps suitable for patch editing. Together, these components form a fully differentiable pipeline in PyTorch3D that links texture modification to model outputs, enabling efficient optimization of patches that remain effective across many viewpoints. The complete process, from image capture to patch printing and physical evaluation, can be completed within a few hours. We demonstrate the effectiveness of the resulting patches through attacks on the YOLOv8 object detection model and discuss remaining challenges and opportunities for improving robustness and scalability.

Singhvi, Vivaan [ORNL] (ORCID:0009000586288221)↗
DOE OSTI2026

Model-independent measurement of the Higgs boson associated production with two jets and decaying to a pair of W bosons in proton-proton collisions at $\sqrt{s}=13$ TeV

A model-independent measurement of the differential production cross section of the Higgs boson decaying into a pair of W bosons, with a final state including two jets produced in association, is presented. In the analysis, events are selected in which the decay products of the two W bosons consist of an electron, a muon, and missing transverse momentum. The model independence of the measurement is maximized by employing a discriminating variable, developed through machine learning, that is agnostic to the signal hypothesis. The analysis is based on proton-proton collision data at $\sqrt{s}=13$ TeV collected with the CMS detector from 2016–2018, corresponding to an integrated luminosity of 138 fb −1 . The production cross section is measured as a function of the difference in azimuthal angle between the two jets. The differential cross section measurements are used to constrain Higgs boson couplings within the standard model effective field theory framework.

Hadron-Hadron Scattering↗
DOE OSTI2022

Identification of driver genes for critical forms of COVID-19 in a deeply phenotyped young patient cohort

The drivers of critical coronavirus disease 2019 (COVID-19) remain unknown. Given major confounding factors such as age and comorbidities, true mediators of this condition have remained elusive. We used a multi-omics analysis combined with artificial intelligence in a young patient cohort where major comorbidities were excluded at the onset. The cohort included 47 “critical” (in the intensive care unit under mechanical ventilation) and 25 “non-critical” (in a non-critical care ward) patients with COVID-19 and 22 healthy individuals. The analyses included whole-genome sequencing, whole-blood RNA sequencing, plasma and blood mononuclear cell proteomics, cytokine profiling, and high-throughput immunophenotyping. An ensemble of machine learning, deep learning, quantum annealing, and structural causal modeling were used. Patients with critical COVID-19 were characterized by exacerbated inflammation, perturbed lymphoid and myeloid compartments, increased coagulation, and viral cell biology. Among differentially expressed genes, we observed up-regulation of the metalloprotease ADAM9. This gene signature was validated in a second independent cohort of 81 critical and 73 recovered patients with COVID-19 and was further confirmed at the transcriptional and protein level and by proteolytic activity. Ex vivo ADAM9 inhibition decreased severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) uptake and replication in human lung epithelial cells. In conclusion, within a young, otherwise healthy, cohort of individuals with COVID-19, we provide the landscape of biological perturbations in vivo where a unique gene signature differentiated critical from non-critical patients. We further identified ADAM9 as a driver of disease severity and a candidate therapeutic target.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2021

Train Like a (Var)Pro: Efficient Training of Neural Networks with Variable Projection

Deep neural networks (DNNs) have achieved state-of-the-art performance across a variety of traditional machine learning tasks, e.g., speech recognition, image classification, and segmentation. The ability of DNNs to efficiently approximate high-dimensional functions has also motivated their use in scientific applications, e.g., to solve partial differential equations and to generate surrogate models. In this paper, we consider the supervised training of DNNs, which arises in many of the above applications. We focus on the central problem of optimizing the weights of the given DNN such that it accurately approximates the relation between observed input and target data. Devising effective solvers for this optimization problem is notoriously challenging due to the large number of weights, nonconvexity, data sparsity, and nontrivial choice of hyperparameters. To solve the optimization problem more efficiently, we propose the use of variable projection (VarPro), a method originally designed for separable nonlinear least-squares problems. Our main contribution is the Gauss--Newton VarPro method (GNvpro) that extends the reach of the VarPro idea to nonquadratic objective functions, most notably cross-entropy loss functions arising in classification. These extensions make GNvpro applicable to all training problems that involve a DNN whose last layer is an affine mapping, which is common in many state-of-the-art architectures. In our four numerical experiments from surrogate modeling, segmentation, and classification, GNvpro solves the optimization problem more efficiently than commonly used stochastic gradient descent (SGD) schemes. Finally, GNvpro finds solutions that generalize well, and in all but one example better than well-tuned SGD methods, to unseen data points.

97 MATHEMATICS AND COMPUTING↗