Decomposing Spiking Neural Networks with Graphical Neural Activity Threads
Explore the source record for details and available documents.
Theilman, Bradley Harold [Sandia National Laborato↗
SEARCH · Search NASA
Results for “Neural network”
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.
Efficiently modeling neural networks on massively parallel computers
Neural networks are a very useful tool for analyzing and modeling complex real world systems. Applying neural network simulations to real world problems generally involves large amounts of data and massive amounts of computation. To efficiently handle the computational requirements of large problems, we have implemented at Los Alamos a highly efficient neural network compiler for serial computers, vector computers, vector parallel computers, and fine grain SIMD computers such as the CM-2 connection machine. This paper describes the mapping used by the compiler to implement feed-forward backpropagation neural networks for a SIMD (Single Instruction Multiple Data) architecture parallel computer. Thinking Machines Corporation has benchmarked our code at 1.3 billion interconnects per second (approximately 3 gigaflops) on a 64,000 processor CM-2 connection machine (Singer 1990). This mapping is applicable to other SIMD computers and can be implemented on MIMD computers such as the CM-5 connection machine. Our mapping has virtually no communications overhead with the exception of the communications required for a global summation across the processors (which has a sub-linear runtime growth on the order of O(log(number of processors)). We can efficiently model very large neural networks which have many neurons and interconnects and our mapping can extend to arbitrarily large networks (within memory limitations) by merging the memory space of separate processors with fast adjacent processor interprocessor communications. This paper will consider the simulation of only feed forward neural network although this method is extendable to recurrent networks.
Farber, Robert M.↗
Optimum Design of Aerospace Structural Components Using Neural Networks
The application of artificial neural networks to capture structural design expertise is demonstrated. The principal advantage of a trained neural network is that it requires a trivial computational effort to produce an acceptable new design. For the class of problems addressed, the development of a conventional expert system would be extremely difficult. In the present effort, a structural optimization code with multiple nonlinear programming algorithms and an artificial neural network code NETS were used. A set of optimum designs for a ring and two aircraft wings for static and dynamic constraints were generated using the optimization codes. The optimum design data were processed to obtain input and output pairs, which were used to develop a trained artificial neural network using the code NETS. Optimum designs for new design conditions were predicted using the trained network. Neural net prediction of optimum designs was found to be satisfactory for the majority of the output design parameters. However, results from the present study indicate that caution must be exercised to ensure that all design variables are within selected error bounds.
Berke, L.↗
A More Accurate Characterization of UH-60A Pitch Link Loads Using Neural Networks
A more accurate, neural-network-based characterization of the full-scale UH-60A maximum, vibratory pitch link loads (MXVPLL) was obtained. The MXVPLL data were taken from the NASA/Army UH-60A Airloads Program flight test database. This database includes data from level flights, and both simple and "complex" maneuvers. In the present context, a complex maneuver was defined as one which involved simultaneous, non-zero aircraft angle-of-bank (associated with turns) and aircraft pitch-rate (associated with a pull-up or a push-over). The present approach combines physical insight followed by the neural networks application. Since existing load factors do not represent the above-defined complex maneuver, a new, combined load factor ('p resent-load-factor') was introduced. A back-propagation type of neural network with five inputs and one output was used to characterize the UH-60A MXVPLL. The neural network inputs were as follows: rotor advance ratio, aircraft gross weight, rotor RPM, air density ratio, and the present-load-factor. The neural network output was the maximum, vibratory pitch link load (MXVPLL). It was shown that a more accurate characterization of the full-scale flight test pitch link loads can be obtained by combining physical insight with a neural-network-based approach.
Kottapalli, Sesi↗
Sea ice classification using fast learning neural networks
A first learning neural network approach to the classification of sea ice is presented. The fast learning (FL) neural network and a multilayer perceptron (MLP) trained with backpropagation learning (BP network) were tested on simulated data sets based on the known dominant scattering characteristics of the target class. Four classes were used in the data simulation: open water, thick lossy saline ice, thin saline ice, and multiyear ice. The BP network was unable to consistently converge to less than 25 percent error while the FL method yielded an average error of approximately 1 percent on the first iteration of training. The fast learning method presented can significantly reduce the CPU time necessary to train a neural network as well as consistently yield higher classification accuracy than BP networks.
Dawson, M. S.↗
Using Neural Networks to Describe Tracer Correlations
Neural networks are ideally suited to describe the spatial and temporal dependence of tracer-tracer correlations. The neural network performs well even in regions where the correlations are less compact and normally a family of correlation curves would be required. For example, the CH4-N2O correlation can be well described using a neural network trained with the latitude, pressure, time of year, and CH4 volume mixing ratio (v.m.r.). In this study a neural network using Quickprop learning and one hidden layer with eight nodes was able to reproduce the CH4-N2O correlation with a correlation co- efficient of 0.9995. Such an accurate representation of tracer-tracer correlations allows more use to be made of long-term datasets to constrain chemical models. Such as the dataset from the Halogen Occultation Experiment (HALOE) which has continuously observed CH4, (but not N2O) from 1991 till the present. The neural network Fortran code used is available for download.
Lary, D. J.↗
Neural-Network Simulator
F77NNS (FORTRAN 77 Neural Network Simulator) computer program simulates popular back-error-propagation neural network. Designed to take advantage of vectorization when used on computers having this capability, also used on any computer equipped with ANSI-77 FORTRAN Compiler. Problems involving matching of patterns or mathematical modeling of systems fit class of problems F77NNS designed to solve. Program has restart capability so neural network solved in stages suitable to user's resources and desires. Enables user to customize patterns of connections between layers of network. Size of neural network F77NNS applied to limited only by amount of random-access memory available to user.
Mitchell, Paul H.↗
Improving the Unsteady Aerodynamic Performance of Transonic Turbines using Neural Networks
A recently developed neural net-based aerodynamic design procedure is used in the redesign of a transonic turbine stage to improve its unsteady aerodynamic performance. The redesign procedure used incorporates the advantages of both traditional response surface methodology and neural networks by employing a strategy called parameter-based partitioning of the design space. Starting from the reference design, a sequence of response surfaces based on both neural networks and polynomial fits are constructed to traverse the design space in search of an optimal solution that exhibits improved unsteady performance. The procedure combines the power of neural networks and the economy of low-order polynomials (in terms of number of simulations required and network training requirements). A time-accurate, two-dimensional, Navier-Stokes solver is used to evaluate the various intermediate designs and provide inputs to the optimization procedure. The procedure yielded a modified design that improves the aerodynamic performance through small changes to the reference design geometry. These results demonstrate the capabilities of the neural net-based design procedure, and also show the advantages of including high-fidelity unsteady simulations that capture the relevant flow physics in the design optimization process.
Rai, Man Mohan↗
Retrieving Aerosol Optical Depth and High Spatial Resolution Ocean Surface Wind Speed from CALIPSO: A Neural Network Approach
An innovative neural network retrieval algorithm is developed for retrieving ocean surface wind speed from the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) lidar measurements. The neural network is trained with CALIPSO ocean surface and atmospheric backscatter measurements together with collocated Advanced Microwave Scanning Radiometer for EOS(AMSR-E) ocean surface wind speed. Ocean surface wind speeds are derived by applying the neural network algorithm to CALIPSO measurements between 2008 and 2020. CALIPSO wind speed measurement of 2015 also compare with Advanced Microwave Scanning Radiometer 2 (AMSR-2) measurements on the Global Change Observation Mission –Water "Shizuku"(GCOM-W) satellite. Aerosol optical depths are then derived from CALIPSO’s ocean surface backscatter signal and theoretical ocean surface reflectance calculated from CALIPSO wind speed and Cox-Munk wind –surface slope variance relation. This CALIPSO wind speed retrieval technique is an improvement from Hu et al. (2008), as it can be applied to most clear skies with optical depths up to 1.5 without making assumptions about aerosol lidar ratio.
wind aerosols CALIPSO↗
Neural Networks for Flight Control
Neural networks are being developed at NASA Ames Research Center to permit real-time adaptive control of time varying nonlinear systems, enhance the fault-tolerance of mission hardware, and permit online system reconfiguration. In general, the problem of controlling time varying nonlinear systems with unknown structures has not been solved. Adaptive neural control techniques show considerable promise and are being applied to technical challenges including automated docking of spacecraft, dynamic balancing of the space station centrifuge, online reconfiguration of damaged aircraft, and reducing cost of new air and spacecraft designs. Our experiences have shown that neural network algorithms solved certain problems that conventional control methods have been unable to effectively address. These include damage mitigation in nonlinear reconfiguration flight control, early performance estimation of new aircraft designs, compensation for damaged planetary mission hardware by using redundant manipulator capability, and space sensor platform stabilization. This presentation explored these developments in the context of neural network control theory. The discussion began with an overview of why neural control has proven attractive for NASA application domains. The more important issues in control system development were then discussed with references to significant technical advances in the literature. Examples of how these methods have been applied were given, followed by projections of emerging application needs and directions.
Jorgensen, Charles C.↗
Anomaly detection of turbopump vibration in Space Shuttle Main Engine using statistics and neural networks
The statistical and neural networks methods have been applied to investigate the feasibility in detecting anomalies in turbopump vibration of SSME. The anomalies are detected based on the amplitude of peaks of fundamental and harmonic frequencies in the power spectral density. These data are reduced to the proper format from sensor data measured by strain gauges and accelerometers. Both methods are feasible to detect the vibration anomalies. The statistical method requires sufficient data points to establish a reasonable statistical distribution data bank. This method is applicable for on-line operation. The neural networks method also needs to have enough data basis to train the neural networks. The testing procedure can be utilized at any time so long as the characteristics of components remain unchanged.
Lo, C. F.↗
Reducing Wind Tunnel Data Requirements Using Neural Networks
The use of neural networks to minimize the amount of data required to completely define the aerodynamic performance of a wind tunnel model is examined. The accuracy requirements for commercial wind tunnel test data are very severe and are difficult to reproduce using neural networks. For the current work, multiple input, single output networks were trained using a Levenberg-Marquardt algorithm for each of the aerodynamic coefficients. When applied to the aerodynamics of a 55% scale model of a U.S. Air Force/ NASA generic fighter configuration, this scheme provided accurate models of the lift, drag, and pitching-moment coefficients. Using only 50% of the data acquired during, the wind tunnel test, the trained neural network had a predictive accuracy equal to or better than the accuracy of the experimental measurements.
Ross, James C.↗
Biologically-informed excitatory and inhibitory ratio for robust spiking neural network training
Spiking neural networks drawing inspiration from biological constraints of the brain promise an energy-efficient paradigm for artificial intelligence. However, challenges exist in identifying guiding principles to train these networks in a robust fashion. In addition, training becomes an even more difficult problem when incorporating biological constraints of excitatory and inhibitory connections. In this work, we identify several key factors, such as low initial firing rates and diverse inhibitory spiking patterns, that determine the overall ability to train in the context of spiking networks with various ratios of excitatory to inhibitory neurons. The results indicate networks with biologically-realistic excitatory:inhibitory ratios can reliably train at low activity levels and in noisy environments. Additionally, the Van Rossum distance, a measure of spike train synchrony, provides insight into the importance of inhibitory neurons to increase network robustness to noise. This work supports further biologically-informed large-scale networks and energy efficient hardware implementations.
bio-inspired computing↗
E-PINNs: Epistemic Physics-Informed Neural Networks
Physics-informed neural networks (PINNs) have demonstrated promise as a framework for solving forward and inverse problems involving partial differential equations. Despite recent progress in the field, it remains challenging to quantify uncertainty in these networks. While techniques such as Bayesian PINNs (B-PINNs) provide a principled approach to capturing epistemic uncertainty through Bayesian inference, they can be computationally expensive for large-scale applications. In this work, we propose Epistemic Physics-Informed Neural Networks (E-PINNs), a framework that uses a small network, the epinet, to efficiently quantify epistemic uncertainty in PINNs. The proposed approach works as an add-on to existing, pre-trained PINNs with a small computational overhead. We demonstrate the applicability of the proposed framework in various test cases and compare the results with B-PINNs using Hamiltonian Monte Carlo (HMC) posterior estimation and dropout-equipped PINNs (Dropout-PINNs). In our experiments, E-PINNs achieve calibrated coverage with competitive sharpness at substantially lower cost. We demonstrate that when B-PINNs produce narrower bands, they under-cover in our tests. E-PINNs also show better calibration than Dropout-PINNs in these examples, indicating a favorable accuracy-efficiency trade-off.
AI for Science↗
Low Temperature Performance of High-Speed Neural Network Circuits
Artificial neural networks, derived from their biological counterparts, offer a new and enabling computing paradigm specially suitable for such tasks as image and signal processing with feature classification/object recognition, global optimization, and adaptive control. When implemented in fully parallel electronic hardware, it offers orders of magnitude speed advantage. Basic building blocks of the new architecture are the processing elements called neurons implemented as nonlinear operational amplifiers with sigmoidal transfer function, interconnected through weighted connections called synapses implemented using circuitry for weight storage and multiply functions either in an analog, digital, or hybrid scheme.
artificial neural networks image processing signal↗
Mesh-based super-resolution of fluid flows with multiscale graph neural networks
A graph neural network (GNN) approach is introduced in this work which enables mesh-based three-dimensional super-resolution of fluid flows. In this framework, the GNN is designed to operate not on the full mesh-based field at once, but on localized meshes of elements (or cells) directly. To facilitate mesh-based GNN representations in a manner similar to spectral (or finite) element discretizations, a baseline GNN layer (termed a message passing layer, which updates local node properties) is modified to account for synchronization of coincident graph nodes, rendering compatibility with commonly used element-based mesh connectivities. Furthermore, the architecture is multiscale in nature, and is comprised of a combination of coarse-scale and fine-scale message passing layer sequences (termed processors) separated by a graph unpooling layer. The coarse-scale processor embeds a query element (alongside a set number of neighboring coarse elements) into a single latent graph representation using coarse-scale synchronized message passing over the element neighborhood, and the fine-scale processor leverages additional message passing operations on this latent graph to correct for interpolation errors. Demonstration studies are performed using hexahedral mesh-based data from Taylor–Green Vortex and backward-facing step flow simulations at Reynolds numbers of 1600 and 3200. Through analysis of both global and local errors, the results ultimately show how the GNN is able to produce accurate super-resolved fields compared to targets in both coarse-scale and multiscale model configurations. Reconstruction errors for fixed architectures were found to increase in proportion to the Reynolds number. Geometry extrapolation studies on a separate cavity flow configuration show promising cross-mesh capabilities of the super-resolution strategy.
Backward-facing step↗
One-shot gas detection with transformer paired neural networks in Mako collected longwave infrared hyperspectral imagery
Explore the source record for details and available documents.
Benham, Kevin Andrew↗
Optimizing the optimizer for physics-informed neural networks and Kolmogorov-Arnold networks
Physics-Informed Neural Networks (PINNs) have revolutionized the computation of PDE solutions by integrating partial differential equations (PDEs) into the neural network’s training process as soft constraints, becoming an important component of the scientific machine learning (SciML) ecosystem. More recently, physics-informed Kolmogorv-Arnold networks (PIKANs) have also shown to be effective and comparable in accuracy with PINNs. In their current implementation, both PINNs and PIKANs are mainly optimized using first-order methods like Adam, as well as quasi-Newton methods such as BFGS and its low-memory variant, L-BFGS. However, these optimizers often struggle with highly nonlinear and non-convex loss landscapes, leading to challenges such as slow convergence, local minima entrapment, and (non)degenerate saddle points. In this study, we investigate the performance of Self- Scaled BFGS (SSBFGS), Self-Scaled Broyden (SSBroyden) methods and other advanced quasi-Newton schemes, including BFGS and L-BFGS with different line search strategies. These methods dynamically rescale updates based on historical gradient information, thus enhancing training efficiency and accuracy. We systematically compare these optimizers – using both PINNs and PIKANs – on key challenging PDEs, including the Burgers, Allen-Cahn, Kuramoto-Sivashinsky, Ginzburg-Landau, and Stokes equations. Additionally, we evaluate the performance of SSBFGS and SSBroyden for Deep Operator Network (DeepONet) architectures, demonstrating their effectiveness for data-driven operator learning. Our findings provide state-of-the-art results with orders-of-magnitude accuracy improvements without the use of adaptive weights or any other enhancements typically employed in PINNs. More broadly, our work reveal insights into the effectiveness of quasi-Newton optimization strategies in significantly improving the convergence and accurate generalization of PINNs and PIKANs.
97 MATHEMATICS AND COMPUTING↗