Search NASA⌕ Search

SEARCH · Search NASA

Results for “DEEP 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 OSTI2025

Applying deep learning methods to develop new models of molecular charge transfer, nonadiabatic dynamics, and nonlinear spectroscopy in the condensed phase

Photon- and field-induced charge transfer has central importance in the generation and storage of electricity, the novel properties of materials, photo-induced catalysis, and electro-optic activity (e.g., photovoltaic cells, fuel cells, and organic chromophores for use in optical fibers and light-emission diodes). These non-equilibrium electronic and chemical transformations are probed by ultrafast, nonlinear spectroscopies. Accurate simulations play a crucial role in our ability to understand, optimize, and control these transformations. This project applies modern deep learning and machine learning (ML) methods to dramatically improve models of electronic dynamics, electronic-nuclear dynamics, and spectroscopic measurements for improved simulations of chemistry in complex environments, far from equilibrium phenomena, and processes in extreme environments, such as materials exposed to strong or resonant fields. This project develops accurate neural net models that go beyond predictive capability to also provide new insight into the fundamental physics underlying electron and nuclear dynamics. To achieve its objectives, this project explores and develops customized versions of high-capacity deep learning algorithms/models. These techniques are developed with an emphasis on fundamental chemical insight, not just predictive accuracy, to assist the development of the next generation of quantum simulation methods.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2021

Global field reconstruction from sparse sensors with Veronoi tessellation-assisted deep learning

Achieving accurate and robust global situational awareness of a complex time-evolving field from a limited number of sensors has been a longstanding challenge. This reconstruction problem is especially difficult when sensors are sparsely positioned in a seemingly random or unorganized manner, which is often encountered in a range of scientific and engineering problems. Moreover, these sensors can be in motion and can become online or offline over time. The key leverage in addressing this scientific issue is the wealth of data accumulated from the sensors. As a solution to this problem, we propose a data-driven spatial field recovery technique founded on a structured grid-based deep-learning approach for arbitrary positioned sensors of any numbers. It should be noted that the naïve use of machine learning becomes prohibitively expensive for global field reconstruction and is furthermore not adaptable to an arbitrary number of sensors. In the present work, we consider the use of Voronoi tessellation to obtain a structured-grid representation from sensor locations enabling the computationally tractable use of convolutional neural networks. One of the central features of the present method is its compatibility with deep-learning based super-resolution reconstruction techniques for structured sensor data that are established for image processing. The proposed reconstruction technique is demonstrated for unsteady wake flow, geophysical data, and three-dimensional turbulence. The current framework is able to handle an arbitrary number of moving sensors, and thereby overcomes a major limitation with existing reconstruction methods. The presented technique opens a new pathway towards the practical use of neural networks for real-time global field estimation.

Fukami, Kai↗
DOE OSTI2025

The Second Skin: A Wearable Sensor Suite that Enables Real-Time Human Biomechanics Tracking Through Deep Learning

Objective: Real-time determination of human kinematics and kinetics could advance biomechanics research and enable valuable applications of biofeedback and generalizable exoskeleton control. Here, this work aims to investigate a taskindependent, user-independent method for obtaining precise realtime joint state estimation across lower-body joints during a wide variety of tasks. Methods: We developed a generalizable sensing approach using a suit comprised of inertial measurement units (IMUs) and pressure insoles. With the suit, we collected a dataset of 33 tasks commonly performed during construction and hazardous waste cleanup (N = 10). We then trained deep learning user-independent, task-agnostic models to estimate joint lowerbody kinematics and dynamics using only worn sensor data. We likewise computed joint kinematics and dynamics analytically from sensor data to serve as a comparison tool for model results. Results: Our models achieved overall angle estimation root-meansquared-errors (RMSE) of 6.56±.92°, 8.60±1.01°, 7.58±.89°, and 6.00±.73° compared to 13.9±.1.3°, 15.31±1.0°, 10.76±.70°, and 7.56±.48° via analytical methods at the lower back, hip, knee, and ankle, respectively. Likewise, our models achieved overall normalized moment estimation RMSEs of .207±.069 Nm/kg, .242±.044 Nm/kg, .202±.038 Nm/kg, and .193±.034 Nm/kg compared to .306±.036 Nm/kg, .407±.021 Nm/kg, 1.18 ±.022 Nm/kg, and 1.73±.071 Nm/kg via analytical methods at the lower back, hip, knee, and ankle, respectively. Conclusion: These results are comparable to other state-of-the-art wearable sensing systems, establishing deep learning as a viable sensing approach that generalizes to new users and tasks. Significance: This work shows promise for enabling accurate real-world biomechanical data collection and enhancement of biofeedback systems and wearable robot control.

Casey, Ryan T. F. [Georgia Institute of Technology↗
DOE OSTI2022

Automatic Detection and Classification of Radio Galaxy Images by Deep Learning

Abstract Surveys conducted by radio astronomy observatories, such as SKA, MeerKAT, Very Large Array, and ASKAP, have generated massive astronomical images containing radio galaxies (RGs). This generation of massive RG images has imposed strict requirements on the detection and classification of RGs and makes manual classification and detection increasingly difficult, even impossible. Rapid classification and detection of images of different types of RGs help astronomers make full use of the observed astronomical image data for further processing and analysis. The classification of FRI and FRII is relatively easy, and there are more studies and literature on them at present, but FR0 and FRI are similar, so it is difficult to distinguish them. It poses a greater challenge to image processing. At present, deep learning has made breakthrough progress in the field of image analysis and processing and has preliminary applications in astronomical data processing. Compared with classification algorithms that can only classify galaxies, object detection algorithms that can locate and classify RGs simultaneously are preferred. In target detection algorithms, YOLOv5 has outstanding advantages in the classification and positioning of small targets. Therefore, we propose a deep-learning method based on an improved YOLOv5 object detection model that makes full use of multisource data, combining FIRST radio with SDSS optical image data, and realizes the automatic detection of FR0, FRI, and FRII RGs. The innovation of our work is that on the basis of the original YOLOv5 object detection model, we introduce the SE Net attention mechanism, increase the number of preset anchors, adjust the network structure of the feature pyramid, and modify the network structure, thereby allowing our model to demonstrate galaxy classification and position detection effects. Our improved model produces satisfactory results, as evidenced by experiments. Overall, the mean average precision (mAP@0.5) of our improved model on the test set reaches 89.4%, which can determine the position (R.A. and decl.) and automatically detect and classify FR0s, FRIs, and FRIIs. Our work contributes to astronomy because it allows astronomers to locate FR0, FRI, and FRII galaxies in a relatively short time and can be further combined with other astronomically generated data to study the properties of these galaxies. The target detection model can also help astronomers find FR0s, FRIs, and FRIIs in future surveys and build a large-scale star RG catalog. Moreover, our work is also useful for the detection of other types of galaxies.

Astronomy & Astrophysics↗
DOE OSTI2024

Predicting biomass comminution: Physical experiment, population balance model, and deep learning

An extended population balance model (PBM) and a deep learning-based enhanced deep neural operator (DNO+) model are introduced for predicting particle size distribution (PSD) of comminuted biomass through a large knife mill. Experimental tests using corn stalks with varied moisture contents, mill blade speeds, and discharge screen sizes are conducted to support model development. A novel mechanism in the extended PBM allows for including additional input parameters such as moisture content, which is not possible in the original PBM. The DNO+ model can include influencing factors of different data types such as moisture content and discharge screen size, which significantly extends the engineering applicability of the standard DNO model that only admits feed PSD and outcome PSD. Test results show that both models are remarkably accurate in the calibration or training parameter space and can be used as surrogate models to provide effective guidance for biomass preprocessing design.

09 BIOMASS FUELS↗
DOE OSTI2021

Deep Learning Based Approach to Integrate MyShake's Trigger Data with ShakeAlert for Faster and Robust EEW Alerts

Earthquake Early Warning (EEW) systems depend on the dense seismic network to make fast and accurate earthquake detections to issue useful early alerts. The recent development and implementation of the ShakeAlert system is relying on the west coast seismic network to detect and estimate the earthquake parameters in real-time. While working on making improvements on the current system, one potential approach is to include more seismic recordings from various sources, such as the low-cost smartphone seismic network, such as the MyShake network, or the Google Android smartphones, since these smartphone seismic networks have much more portable stations (phones) to potentially provide useful data to the system. This report summarizes the initial exploration of using a deep learning approach to combine the data from both traditional seismic stations and the smartphone data in California. Due to the heterogeneity nature of the data, we aggregate data from the phones (using simulation data), as well as that from traditional seismic stations to grid cells. By generating real-time triggering ratio grid cell maps, the designed deep learning algorithm can process the data from multiple sources and detect the earthquake faster than only using that from a traditional seismic network.

58 GEOSCIENCES↗
DOE OSTI2022

Matrix-Model Simulations Using Quantum Computing, Deep Learning, and Lattice Monte Carlo

Matrix quantum mechanics plays various important roles in theoretical physics, such as a holographic description of quantum black holes, and it underpins the only practical numerical approach to the study of complex high-dimensional supergravity theories. Understanding quantum black holes and the role of entanglement in a holographic setup is of paramount importance for the realization of a quantum theory of gravity. Moreover, a complete numerical understanding of the holographic duality and the emergence of geometric space-time features from microscopic degrees of freedom could pave the way for new discoveries in quantum information science. Euclidean lattice Monte Carlo simulations are the de facto numerical tool for understanding the spectrum of large matrix models and have been used to test the holographic duality. However, they are not tailored to extract dynamical properties or even the quantum wave function of the ground state of matrix models. Quantum computing and deep learning provide potentially useful approaches to study the dynamics of matrix quantum mechanics. If successful in the context of matrix models, these rapidly improving numerical techniques could become the new Swiss army knife of quantum gravity practitioners. In this paper, we perform the first systematic survey for quantum computing and deep-learning approaches to matrix quantum mechanics, comparing them to lattice Monte Carlo simulations. These provide baseline benchmarks before addressing more complicated problems. In particular, we test the performance of each method by calculating the low-energy spectrum.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗
DOE OSTI2021

Deep-learning-aided forward optical coherence tomography endoscope for percutaneous nephrostomy guidance

Percutaneous renal access is the critical initial step in many medical settings. In order to obtain the best surgical outcome with minimum patient morbidity, an improved method for access to the renal calyx is needed. In our study, we built a forward-view optical coherence tomography (OCT) endoscopic system for percutaneous nephrostomy (PCN) guidance. Porcine kidneys were imaged in our experiment to demonstrate the feasibility of the imaging system. Three tissue types of porcine kidneys (renal cortex, medulla, and calyx) can be clearly distinguished due to the morphological and tissue differences from the OCT endoscopic images. To further improve the guidance efficacy and reduce the learning burden of the clinical doctors, a deep-learning-based computer aided diagnosis platform was developed to automatically classify the OCT images by the renal tissue types. Convolutional neural networks (CNN) were developed with labeled OCT images based on the ResNet34, MobileNetv2 and ResNet50 architectures. Nested cross-validation and testing was used to benchmark the classification performance with uncertainty quantification over 10 kidneys, which demonstrated robust performance over substantial biological variability among kidneys. ResNet50-based CNN models achieved an average classification accuracy of 82.6%±3.0%. The classification precisions were 79%±4% for cortex, 85%±6% for medulla, and 91%±5% for calyx and the classification recalls were 68%±11% for cortex, 91%±4% for medulla, and 89%±3% for calyx. Interpretation of the CNN predictions showed the discriminative characteristics in the OCT images of the three renal tissue types. The results validated the technical feasibility of using this novel imaging platform to automatically recognize the images of renal tissue structures ahead of the PCN needle in PCN surgery.

Wang, Chen↗
DOE OSTI2021

Inpainting radar missing data regions with deep learning

Abstract. Missing and low-quality data regions are a frequent problem for weather radars. They stem from a variety of sources: beam blockage, instrument failure, near-ground blind zones, and many others. Filling in missing data regions is often useful for estimating local atmospheric properties and the application of high-level data processing schemes without the need for preprocessing and error-handling steps – feature detection and tracking, for instance. Interpolation schemes are typically used for this task, though they tend to produce unrealistically spatially smoothed results that are not representative of the atmospheric turbulence and variability that are usually resolved by weather radars. Recently, generative adversarial networks (GANs) have achieved impressive results in the area of photo inpainting. Here, they are demonstrated as a tool for infilling radar missing data regions. These neural networks are capable of extending large-scale cloud and precipitation features that border missing data regions into the regions while hallucinating plausible small-scale variability. In other words, they can inpaint missing data with accurate large-scale features and plausible local small-scale features. This method is demonstrated on a scanning C-band and vertically pointing Ka-band radar that were deployed as part of the Cloud Aerosol and Complex Terrain Interactions (CACTI) field campaign. Three missing data scenarios are explored: infilling low-level blind zones and short outage periods for the Ka-band radar and infilling beam blockage areas for the C-band radar. Two deep-learning-based approaches are tested, a convolutional neural network (CNN) and a GAN that optimize pixel-level error or combined pixel-level error and adversarial loss respectively. Both deep-learning approaches significantly outperform traditional inpainting schemes under several pixel-level and perceptual quality metrics.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2020

Deeply uncertain: comparing methods of uncertainty quantification in deep learning algorithms

We present a comparison of methods for uncertainty quantification (UQ) in deep learning algorithms in the context of a simple physical system. Three of the most common uncertainty quantification methods - Bayesian Neural Networks (BNN), Concrete Dropout (CD), and Deep Ensembles (DE) - are compared to the standard analytic error propagation. We discuss this comparison in terms endemic to both machine learning ("epistemic" and "aleatoric") and the physical sciences ("statistical" and "systematic"). The comparisons are presented in terms of simulated experimental measurements of a single pendulum - a prototypical physical system for studying measurement and analysis techniques. Our results highlight some pitfalls that may occur when using these UQ methods. For example, when the variation of noise in the training set is small, all methods predicted the same relative uncertainty independently of the inputs. This issue is particularly hard to avoid in BNN. On the other hand, when the test set contains samples far from the training distribution, we found that no methods sufficiently increased the uncertainties associated to their predictions. This problem was particularly clear for CD. In light of these results, we make some recommendations for usage and interpretation of UQ methods.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2025

Multi-modality deep learning for pulse prediction in homogeneous nonlinear systems via parametric conversion

In this Letter, we introduce FusionNet, a multi-modality deep learning framework designed to predict and analyze output pulses in high-power rare-earth-doped laser systems driving parametric conversion in homogeneous guided nonlinear media. FusionNet integrates temporal, spectral, and physical experimental conditions to model ultrafast nonlinear phenomena, including parametric nonlinear frequency conversion, self-phase modulation, and cross-phase modulation in homogeneous guided systems such as gas-filled hollow-core fibers. These systems bridge physical models with experimental data, advancing our understanding of light-guiding principles and nonlinear interactions while expediting the design and optimization of on-demand high-power, high-brightness systems. Our results demonstrate a 73% reduction in prediction error and an 83% improvement in computational efficiency compared to conventional neural networks. This work establishes a new paradigm for accelerating parametric simulations and optimizing experimental designs in high-power laser systems, with further implications for high-precision spectroscopy, quantum information science, and distributed entangled interconnects.

47 OTHER INSTRUMENTATION↗
DOE OSTI2024

Datasets and U-Net Model for "A Deep Learning Based Framework to Identify Undocumented Orphaned Oil and Gas Wells from Historical Maps: a Case Study for California and Oklahoma"

This dataset has results and the model associated with the publication Ciulla et al., (2024). It contains a U-Net semantic segmentation model (unet_model.h5) and associated code implemented in tensorflow 2.0 for the model training and identification of oil and gas well symbols in USGS historical topographic maps (HTMC). Given a quadrangle map (7.5 minutes), downloadable at this url: https://ngmdb.usgs.gov/topoview/, and a list of coordinates of the documented wells present in the area, the model returns the coordinates of oil and gas symbols in the HTMC maps. For reproducibility of our workflow, we provide a sample map in California and the documented well locations for the entire State of California (CalGEM_AllWells_20231128.csv) downloaded from https://www.conservation.ca.gov/calgem/maps/Pages/GISMapping2.aspx. Additionally, the locations of 1,301 potential undocumented orphaned wells identified using our deep learning framework or the counties of Los Angeles and Kern in California, and Osage and Oklahoma in Oklahoma are provided in the file found_potential_UOWs.zip. The results of the visual inspection of satellite imagery in Osage County is in the file visible_potential_UOWs.zip. The dataset also includes a custom tool to validate the detected symbols in the HTMC maps (vetting_tool.py). More details about the methodology can be found in the associated paper: Ciulla, F., Santos, A., Jordan, P., Kneafsey, T., Biraud, S.C., and Varadharajan, C. (2024) A Deep Learning Based Framework to Identify Undocumented Orphaned Oil and Gas Wells from Historical Maps: a Case Study for California and Oklahoma. Accepted for publication in Environmental Science and Technology. The geographical coordinates provided correspond to the locations of potential undocumented orphaned oil and gas wells (UOWs) extracted from historical maps. The actual presence of wells need to be confirmed with on-the-ground investigations. For your safety, do not attempt to visit or investigate these sites without appropriate safety training, proper equipment, and authorization from local authorities. Approaching these well sites without proper personal protective equipment (PPE) may pose significant health and safety risks. Oil and gas wells can emit hazardous gasses including methane, which is flammable, odorless and colorless, as well as hydrogen sulfide, which can be fatal even at low concentrations. Additionally, there may be unstable ground near the wellhead that may collapse around the wellbore. This dataset was prepared as an account of work sponsored by the United States Government. While this document is believed to contain correct information, neither the United States Government nor any agency thereof, nor the Regents of the University of California, nor any of their employees, makes any warranty, express or implied, or assumes any legal responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by its trade name, trademark, manufacturer, or otherwise, does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States Government or any agency thereof, or the Regents of the University of California. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof or the Regents of the University of California.

Artificial Intelligence↗
DOE OSTI2021

Extracting low energy signals from raw LArTPC waveforms using deep learning techniques -- A proof of concept

We investigate the feasibility of using deep learning techniques, in the form of a one-dimensional convolutional neural network (1D-CNN), for the extraction of signals from the raw waveforms produced by the individual channels of liquid argon time projection chamber (LArTPC) detectors. A minimal generic LArTPC detector model is developed to generate realistic noise and signal waveforms used to train and test the 1D-CNN, and evaluate its performance on low-level signals. We demonstrate that our approach overcomes the inherent shortcomings of traditional cut-based methods by extending sensitivity to signals with ADC values below their imposed thresholds. This approach exhibits great promise in enhancing the capabilities of future generation neutrino experiments like DUNE to carry out their low-energy neutrino physics programs.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗
DOE OSTI2024

DeepUQ: Assessing the Aleatoric Uncertainties from two Deep Learning Methods

Assessing the quality of aleatoric uncertainty estimates from uncertainty quantification (UQ) deep learning methods is important in scientific contexts, where uncertainty is physically meaningful and important to characterize and interpret exactly. We systematically compare aleatoric uncertainty measured by two UQ techniques, Deep Ensembles (DE) and Deep Evidential Regression (DER). Our method focuses on both zero-dimensional (0D) and two-dimensional (2D) data, to explore how the UQ methods function for different data dimensionalities. We investigate uncertainty injected on the input and output variables and include a method to propagate uncertainty in the case of input uncertainty so that we can compare the predicted aleatoric uncertainty to the known values. We experiment with three levels of noise. The aleatoric uncertainty predicted across all models and experiments scales with the injected noise level. However, the predicted uncertainty is miscalibrated to $\rm{std}(\sigma_{\rm al})$ with the true uncertainty for half of the DE experiments and almost all of the DER experiments. The predicted uncertainty is the least accurate for both UQ methods for the 2D input uncertainty experiment and the high-noise level. While these results do not apply to more complex data, they highlight that further research on post-facto calibration for these methods would be beneficial, particularly for high-noise and high-dimensional settings.

Nevin, Rebecca↗
DOE OSTI2023

A robust estimator of mutual information for deep learning interpretability

Abstract We develop the use of mutual information (MI), a well-established metric in information theory, to interpret the inner workings of deep learning (DL) models. To accurately estimate MI from a finite number of samples, we present GMM-MI (pronounced ‘Jimmie’), an algorithm based on Gaussian mixture models that can be applied to both discrete and continuous settings. GMM-MI is computationally efficient, robust to the choice of hyperparameters and provides the uncertainty on the MI estimate due to the finite sample size. We extensively validate GMM-MI on toy data for which the ground truth MI is known, comparing its performance against established MI estimators. We then demonstrate the use of our MI estimator in the context of representation learning, working with synthetic data and physical datasets describing highly non-linear processes. We train DL models to encode high-dimensional data within a meaningful compressed (latent) representation, and use GMM-MI to quantify both the level of disentanglement between the latent variables, and their association with relevant physical quantities, thus unlocking the interpretability of the latent representation. We make GMM-MI publicly available in this GitHub repository.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗
DOE OSTI2023

Identifying common stored product insects using automated deep learning methods

Monitoring stored product insect pests is a common practice for post-harvest management of stored grain and grain-based commodities, which helps ensure product quality from harvest to final consumer. Current methods of sampling and monitoring can be time-consuming, labor-intensive, expensive and require expertise in insect identification. Therefore, this study aims to develop an image-based automated identification system for common stored product insect species using deep-learning methods. Top-down images of the common stored product adult insect species of Rhyzopertha dominica, Cryptolestes ferrugineus, Tribolium castaneum, Sitophilus oryzae, and Oryzaephilus surinamensis were acquired and analyzed. Deep learning-based, state-of-the-art Convolutional Neural Networks (CNN) models (ResNet-50, MobileNet-v2, DarkNet-53, and EfficientNet-b0) were fine-tuned with a transfer learning approach to classify the insect species. All models were able to correctly identify the insect species with at least 96% accuracy and with few misclassifications. One issue with trained CNNs is that they do not explain the reasoning for the classification and are often called a “black box”. Therefore, visualization methods called Gradient-weighted Class Activation Mapping (Grad-CAM) were implemented to explore the black box network. The Grad-CAM uses heat maps to highlight the major image features that the network focused on to make insect species predictions. The Grad-CAM verifies the network's prediction and also helps improve network performance. This study contributes to the overall goal of developing a camera-based system for monitoring stored grain insects. As a result, the developed system would empower warehouse, flour mills, and other food facilities with a tool to quickly and accurately identify insect species in stored product environments and could be implemented as part of a close to real-time monitoring system.

60 APPLIED LIFE SCIENCES↗
DOE OSTI2025

Techno-economic implications and cost of forecasting errors in solar PV power production using optimized deep learning models

Accurate solar Photovoltaic (PV) power forecasting is important for enhancing both the performance and economic feasibility of PV systems. This study evaluates several deep learning models, including Dense Neural Networks (DNN), Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), and a hybrid LSTMCNN model, for predicting PV power production one day in advance. Prior to optimization, the models exhibited relatively high errors, with the best model (DNN) achieving a Root Mean Square Error (RMSE) of 31.13 kW and a coefficient of determination (R 2 ) of 62.15 %. After employing Bayesian optimization, the LSTM-CNN model demonstrated the best performance, with the RMSE reduced to 9.79 kW and R 2 improved to 97.62 %, showcasing significant enhancement in predictive accuracy. Here, the economic evaluation considered three cases: rewards for underestimation (0.08 USD/kWh), no rewards, and penalties for both over-and underestimation (120 % of the utility tariff). In the rewards scenario, the LSTM-CNN model reduced the Levelized Cost of Electricity (LCOE) by 4 %, while in the penalty scenario, a backup diesel generator would have increased the LCOE by 49 %. Additionally, the LSTM-CNN model minimized financial losses, achieving the lowest penalties and maximizing net cash flow compared to other models, demonstrating its overall technical and economic superiority.

Deep learning↗
DOE OSTI2022

Automatic Generation of High-Performance Convolution Kernels on ARM CPUs for Deep Learning

In this work, we present FastConv, a template-based code auto-generation open source library that can automatically generate high-performance deep learning convolution kernels of arbitrary matrices/tensors shapes. FastConv is based on the Winograd algorithm, which is reportedly the highest performing algorithm for the time-consuming convolution layers of convolutional neural networks. ARM CPUs cover a wide range designs and specifications, from embedded devices to HPC-grade CPUs. The leads to the dilemma of how to consistently optimize Winograd-based convolution solvers for convolution layers of different shapes. FastConv addresses this problem by using templates to auto-generate multiple shapes of tuned kernels variants suitable for skinny tall matrices. As a performance portable library, FastConv transparently searches for the best combination of kernel shapes, cache tiles, scheduling of loop orders, packing strategies, access patterns, and online/offline computations. Auto-tuning is used to search the parameter configuration space for the best performance for a given target architecture and problem size. The experiments with layer-wise evaluation on the VGG--16 model confirms a 1.25x performance gains is got by tuning the Winograd library. Integrated comparison results shows 1.02x to 1.40x, 1.14x to 2.17x, and 1.22x and 2.48x speedup is achieved over NNPACK, Arm NN, and FeatherCNN on the Kunpeng 920 beside few cases. Furthermore, problem size performance portability experiments with various convolution shapes shows that FastConv achieves 1.2x to 1.7x speedup and 2x to 22x speedup over NNPACK and ARM NN inference engine using Winograd on Kunpeng 920 . CPU performance portability evaluation on the VGG--16 show an average speedup over NNPACK of 1.42x, 1.21x, 1.26x, 1.37x, 2.26x, and 11.02x is observed on Kunpeng 920, Snapdragon 835, 855, 888, Apple M1, and AWS Graviton2, respectively.

97 MATHEMATICS AND COMPUTING↗