Search NASA⌕ Search

SEARCH · Search NASA

Results for “3D neural networks”

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 271 records · Page 15
All sourcesNASA NTRSDOE OSTI
NASA NTRSDate not supplied

Synthetic Data Generation for 3D Mesh Prediction and Spatial Reasoning During Multi-Agent Robotic Missions

In-space assembly operations require accurate reasoning over the pose, location, and structural organization of both the autonomous agents and assembly materials. In a full six-degree-of-freedom space, an accurate understanding of the full three-dimensional structure of the object of interest greatly enriches information for pose estimation and collision planning. Current methods of predicting pose estimation require a priori understanding of the shape of the object. Additionally, visual information in the space environment is impacted by variations in contrast and illumination. Using synthetic data allows us to rapidly generate large datasets with in varying environments and lighting conditions.This work details the generation of synthetic data used to explore the use of a region-based convolutional neural networks to detect objects of interest and predict a voxel-based three-dimensional mesh in order to understand their full three-dimensional shape. This mesh provides useful spatial information during in-space assembly operations without requiring either the complexity of maintaining models over the progress of building an object or observations from multiple angles. The generated meshes are then compared to that of ground truth in order to measure its performance.

synthetic data↗
DOE OSTI2021

Boost of the Bio-memristor Performance for Artificial Electronic Synapses by Surface Reconstruction

Biomaterial-based memristors (bio-memristors) are often adopted to emulate biological synapse functions and applied to construct neural computing networks in brain-inspired chip systems. However, the randomness of conductive filament formation in bio-memristors inhibits their switching performance by causing the dispersion of the device-switching parameters. In this case, a facile porous silk fibroin (p-SF) memristor was obtained through a protein surface reconstruction strategy, in which the size of the hole can be adjusted by the density of hybrid nanoseeds. Here, the porous SF memristors exhibit greatly enhanced electrical characteristics, including uniform I-V cycles, centralized distribution of the switching voltages, and both high and low resistances, compared to devices without pores. The results of three-dimensional (3D) simulations based on classical density functional theory (cDFT) suggest that the reconstructed pores in the SF layers guide the formation and fracture of Ag filaments under an electric field and enhance the overall conductivity by separating Ag + ion and electron diffusion pathways. Ag + ions are predicted to preferentially diffuse through pores, whereas electrons diffuse through the SF network. Interestingly, the device conductance can be bidirectionally modulated gradually by positive and negative voltages, can faithfully simulate short-term and long-term plasticity, and can even realize the triplet-spike-timing-dependent plasticity (triplet-STDP) rule, which can be used for pattern recognition in biological systems. The simulation results reveal that a memristor network of this type has an accuracy of ~95.78% in memory learning and the capability of pattern learning. This work provides a facile technology route to improve the performance of bionic-material memristors.

36 MATERIALS SCIENCE↗
NASA NTRSDate not supplied

Synthetic Data Generation for 3D Mesh Prediction and Spatial Reasoning During Multi-Agent Robotic Missions

In-space assembly operations require accurate reasoning over the pose, location, and structural organization of both the autonomous agents and assembly materials. In a full six-degree-of-freedom space, an accurate understanding of the full three-dimensional structure of the object of interest greatly enriches information for pose estimation and collision planning. Current methods of predicting pose estimation require a priori understanding of the shape of the object. Additionally, visual information in the space environment is impacted by variations in contrast and illumination. Using synthetic data allows us to rapidly generate large datasets with in varying environments and lighting conditions. This work details the generation of synthetic data used to explore the use of a region-based convolutional neural networks to detect objects of interest and predict a voxel-based three-dimensional mesh in order to understand their full three-dimensional shape. This mesh provides useful spatial information during in-space assembly operations without requiring either the complexity of maintaining models over the progress of building an object or observations from multiple angles. The generated meshes are then compared to that of ground truth in order to measure its performance.

James Ecker↗
DOE OSTI2025

Many-body expansion based machine learning models for octahedral transition metal complexes

Abstract Graph-based machine learning (ML) models for material properties show great potential to accelerate virtual high-throughput screening of large chemical spaces. However, in their simplest forms, graph-based models do not include any 3D information and are unable to distinguish stereoisomers such as those arising from different orderings of ligands around a metal center in coordination complexes. In this work we present a modification to revised autocorrelation descriptors, a molecular graph featurization method, for predicting spin state dependent properties of octahedral transition metal complexes (TMCs). Inspired by analytical semi-empirical models for TMCs, the new modeling strategy is based on the many-body expansion (MBE) and allows one to tune the captured stereoisomer information by changing the truncation order of the MBE. We present the necessary modifications to include this approach in two commonly used ML methods, kernel ridge regression and feed-forward neural networks. On a test set composed of all possible isomers of binary TMCs, the best MBE models achieve mean absolute errors (MAEs) of 2.75 kcal mol −1 on spin-splitting energies and 0.26 eV on frontier orbital energy gaps, a 30%–40% reduction in error compared to models based on our previous approach. We also observe improved generalization to previously unseen ligands where the best-performing models exhibit MAEs of 4.00 kcal mol −1 (i.e. a 0.73 kcal mol −1 reduction) on the spin-splitting energies and 0.53 eV (i.e. a 0.10 eV reduction) on the frontier orbital energy gaps. Because the new approach incorporates insights from electronic structure theory, such as ligand additivity relationships, these models exhibit systematic generalization from homoleptic to heteroleptic complexes, allowing for efficient screening of TMC search spaces.

Meyer, Ralf (ORCID:0000000322360261)↗
DOE OSTI2020

Beam Hardening Artifact Reduction in X-Ray CT Reconstruction of 3D Printed Metal Parts Leveraging Deep Learning and CAD Models

Nondestructive evaluation (NDE) of additively manufactured (AM) parts is important for understanding the impacts of various process parameters and qualifying the built part. X-ray computed tomography (XCT) has played a critical role in rapid NDE and characterization of AM parts. However, XCT of metal AM parts can be challenging because of artifacts produced by standard reconstruction algorithms as a result of a confounding effect called “beam hardening.” Beam hardening artifacts complicate the analysis of XCT images and adversely impact the process of detecting defects, such as pores and cracks, which is key to ensuring the quality of the parts being printed. In this work, we propose a novel framework based on using available computer-aided design (CAD) models for parts to be manufactured, accurate XCT simulations, and a deep-neural network to produce high-quality XCT reconstructions from data that are affected by noise and beam hardening. Using extensive experiments with simulated data sets, we demonstrate that our method can significantly improve the reconstruction quality, thereby enabling better detection of defects compared with the state of the art. We also present promising preliminary results of applying the deep networks trained using CAD models to experimental data obtained from XCT of an AM jet-engine turbine blade.

Ziabari, Amir↗
DOE OSTI2024

Z-Target Radiography Postprocessing With A Deep Convolution Neural Network

Analyzing X-ray radiographs is crucial for understanding target behavior in Inertial Confinement Fusion (ICF) and High Energy Density (HED) platforms. However, the density of Magneto Raleigh Taylor (MRT) bands and limitations of target materials often obscure relevant spike growth and density information. To address this issue, machine learning postprocessing techniques can be applied to remove darkened regions in radiography images. In this study, a novel method is presented for removing MRT darkened regions from z-target radiographs using a convolutional neural network (CNN). The CNN, consisting of six layers, treats the darkened regions as noise and employs a mixed loss function and end-to-end frameworks to suppress them while preserving sharpness. The six-layer architecture is designed to effectively learn features when provided with a larger volume of learning space. Each layer is optimized using a mixed loss function that combines a standard loss pixel approach with a multi-scaled structural similarity index loss, which considers luminance, contrast, and structure in local neighborhoods. This approach is particularly beneficial for capturing the stochastic structure of MRT limbs. Due to the limited availability of experimental data, training is conducted using synthetic target radiography from 3D Alegra simulations.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2022

Real-time X-ray phase-contrast imaging using SPINNet—a speckle-based phase-contrast imaging neural network

X-ray phase-contrast imaging has become indispensable for visualizing samples with low absorption contrast. In this regard, speckle-based techniques have shown significant advantages in spatial resolution, phase sensitivity, and implementation flexibility compared with traditional methods. However, the computational cost associated with data inversion has hindered their wider adoption. By exploiting the power of deep learning, we developed a speckle-based phase-contrast imaging neural network (SPINNet) that significantly improves the imaging quality and boosts the phase retrieval speed by at least 2 orders of magnitude compared to existing methods. To achieve this performance, we combined SPINNet with a coded-mask-based technique, an enhanced version of the speckle-based method. Using this scheme, we demonstrate the simultaneous reconstruction of absorption and phase images on the order of 100 ms, where a traditional correlation-based analysis would take several minutes even with a cluster. In addition to significant improvement in speed, our experimental results show that the imaging and phase retrieval quality of SPINNet outperform existing single-shot speckle-based methods. Furthermore, we successfully demonstrate SPINNet application in x-ray optics metrology and 3D x-ray phase-contrast tomography. Our result shows that SPINNet could enable many applications requiring high-resolution and fast data acquisition and processing, such as in situ and in operando 2D and 3D phase-contrast imaging and real-time at-wavelength metrology and wavefront sensing.

36 MATERIALS SCIENCE↗
NASA NTRS1989

Cloud cover determination in polar regions from satellite imagery

The objectives are to develop a suitable validation data set for evaluating the effectiveness of the International Satellite Cloud Climatology Project (ISCCP) algorithm for cloud retrieval in polar regions, to identify limitations of current procedures and to explore potential means to remedy them using textural classifiers, and to compare synoptic cloud data from model runs with observations. Toward the first goal, a polar data set consisting of visible, thermal, and passive microwave data was developed. The AVHRR and SMMR data were digitally merged to a polar stereographic projection with an effective pixel size of 5 sq km. With this data set, two unconventional methods of classifying the imagery for the analysis of polar clouds and surfaces were examined: one based on fuzzy sets theory and another based on a trained neural network. An algorithm for cloud detection was developed from an early test version of the ISCCP algorithm. This algorithm includes the identification of surface types with passive microwave, then temporal tests at each pixel location in the cloud detection phase. Cloud maps and clear sky radiance composites for 5 day periods are produced. Algorithm testing and validation was done with both actural AVHRR/SMMR data, and simulated imagery. From this point in the algorithm, groups of cloud pixels are examined for their spectral and textural characteristics, and a procedure is developed for the analysis of cloud patterns utilizing albedo, IR temperature, and texture. In a completion of earlier work, empirical analyses of arctic cloud cover were explored through manual interpretations of DMSP imagery and compared to U.S. Air Force 3D-nephanalysis. Comparisons of observed cloudiness from existing climatologies to patterns computed by the GISS climate model were also made.

Barry, R. G.↗
DOE OSTI2022

Bridging microscopy with molecular dynamics and quantum simulations: an atomAI based pipeline

Recent advances in (scanning) transmission electron microscopy have enabled a routine generation of large volumes of high-veracity structural data on 2D and 3D materials, naturally offering the challenge of using these as starting inputs for atomistic simulations. In this fashion, the theory will address experimentally emerging structures, as opposed to the full range of theoretically possible atomic configurations. However, this challenge is highly nontrivial due to the extreme disparity between intrinsic timescales accessible to modern simulations and microscopy, as well as latencies of microscopy and simulations per se. Addressing this issue requires as a first step bridging the instrumental data flow and physics-based simulation environment, to enable the selection of regions of interest and exploring them using physical simulations. Here we report the development of the machine learning workflow that directly bridges the instrument data stream into Python-based molecular dynamics and density functional theory environments using pre-trained neural networks to convert imaging data to physical descriptors. Additionally, the pathways to ensure structural stability and compensate for the observational biases universally present in the data are identified in the workflow. This approach is used for a graphene system to reconstruct optimized geometry and simulate temperature-dependent dynamics including adsorption of Cr as an ad-atom and graphene healing effects. However, it is universal and can be used for other material systems.

36 MATERIALS SCIENCE↗
DOE OSTI2026

Knowledge gaps for neuromorphic ionic computing

BACKGROUND Neuromorphic computing, inspired by the human brain’s ability to process information efficiently, represents a transformative approach to computation. In this Review, we explore the emerging field of neuromorphic ionic computing, which leverages ionic conduction and coupling to mimic neural processes, and identify critical knowledge gaps that must be addressed to realize its full potential. A central theme of the discussion is energy efficiency, a challenge that is both a limitation and an opportunity for this technology. Although complementary metal-oxide semiconductor (CMOS)–based neuromorphic technologies have made strides in scaling to billions of neurons and are increasingly applied in artificial intelligence and numerical computing, they remain orders of magnitude behind the human brain in terms of connectivity and energy efficiency. Neuromorphic ionic computing promises to overcome these limitations by leveraging the distinct architectural and operational principles of the brain. Our brains achieve this energy efficiency by combining several key features: using the same network elements to store and process information; using an incredibly complex and massively interconnected three-dimensional (3D) network of locally active elements that enables sparsity, robustness in the presence of noise, adaptation, and life-long learning; computing at comparatively low voltage and frequency; and last, taking advantage of a plethora of ions and small molecules as information carriers. Here, we propose that ionic computing systems can take advantage of similar features to achieve substantial gains in energy efficiency. ADVANCES Since the first reports of neuromorphic ionic behavior in nanofluidic channels, we have witnessed an explosion of reports that used ionic devices to produce synaptomimetic behaviors. However, achieving the goals of ionic computing requires not only implementation of much more sophisticated device functionality but also overcoming fundamental barriers in materials science, device architecture, and system integration. Current ionic devices, even those incorporating state-of-the-art materials, still suffer from limited functionality and stability, which restrict their performance and increase energy demands. Developing new materials with enhanced ionic properties is essential to overcome these limitations. Similarly, the design of neuromorphic devices must evolve to leverage the particular advantages of ionic processes. Existing architectures often follow a single-information-carrier logic of conventional electronics or are constructed of mesoscale fluidics, failing to capitalize on the energy-efficient mechanisms inherent to ionic systems or implement the multiple-information-carrier paradigm. Current neuromorphic chips focus on large-scale networks of analog memory elements based on mechanisms such as charge trap (flash), filamentary, phase change, or spin, which are built on top of a network of artificial CMOS neurons. Although such prototype networks have achieved impressive performance, it is difficult to envision how they can implement the key features such as massive connectivity, sophisticated plasticity, adaptability, sparsity, and “multichromatic” computing. Although small-scale devices have demonstrated promising results, integrating them, maintaining energy efficiency, and implementing temperature control as systems grow in complexity and size to computationally relevant scale remain major hurdles. Furthermore, interfacing neuromorphic ionic devices with existing computing technologies presents technical and conceptual challenges that will require innovative approaches that combine insights from neuroscience, materials science, and engineering. OUTLOOK Despite these challenges, the potential impact of neuromorphic ionic computing is profound with potential applications ranging from artificial intelligence to robotics and beyond. We also argue that neuromorphic ionic computing systems should not, at least in the beginning, compete with CMOS technologies but rather should focus on applications that require extreme energy efficiency with chemical and/or biological compatibility, such as biomedical applications (for example, brain-computer interfaces), environmental monitoring, and agricultural and food applications. Ultimately, this Review highlights the crucial role of interdisciplinary collaboration in advancing the field. Neuromorphic ionic computing is not merely a technological innovation; it represents a substantial step toward sustainable computation, aligning with the growing demand for energy-conscious solutions in a world that is increasingly reliant on data and computation.

Neuromorphic↗
DOE OSTI2021

Integrating static PRA information with risk informed safety margin characterization (RISMC) simulation methods

The overall objective of the project was to develop a computationally feasible and user-friendly mechanized process to integrate traditional probabilistic risk assessment (PRA) and dynamic PRA (DPRA) results. Starting with the systematic identification of items in an existing PRA that need dynamic augmentation, the project used a generic 4-loop pressurized reactor (PWR) and 3-loop PWR as example plants. Station blackout (SBO) and large break loss of coolant accident SBLOCA) were selected as the example initiating events. Using the traditional event-tree (ET)/fault-tree (FT) methodology augmented by dynamic evet tree approach, the potential consequences of the initiating events were simulated with RELAP-3D and MELCOR/RASCAL codes to cover Level 1 through Level 3 of PRA. RAVEN and ADAPT software were used to generate Level 1 simulations with RELAP-3D and Level 2/3 simulations with MELCOR (Level 2)/RASCAL (Level 3), respectively. Example branching conditions (BCs) for SBO included AC power recovery time, valve repair failure time, reactor coolant pump leak time/break size and emergency power supply duration to a total of 9. Example BCs for LOCA included off-site power recovery time, diesel generator power recovery time, auxiliary feed water system operation time, safety relief valve failure to open upon demand, reactor coolant pump seal break time and size to a total of 21. Each RELAP-3D simulation (9,587 scenarios) was labelled OK or Core Damage based on the maximum allowed peak clad temperature (2,100oF). Each MELCOR simulation (4610 scenarios) was labeled as Bin over 10rem or Bin 0-10rem based on the dose at the site boundary. The scenarios were clustered based on the criteria above using the mean shift methodology. Classical PRA (CPRA) and DPRA results were compared to identify the ET sequences that need DPRA augmentation. Several approaches were proposed for the incorporation of these sequences into CPRA using clustering with the mean shift methodology, restructuring the CPRA ETs by adding new BCs/sequences, and using the concept of a limit surface. Procedures for decision making regarding the possible consequences of an initiating event (e.g. core damage or not, site evacuation or not) were developed using a convolutional neural network (CNN), a recurrent neural network (RNN) and a transformer neural network (TNN). The project has led to two PhD degrees, three archival journal papers and five refereed conference proceedings.

21 SPECIFIC NUCLEAR REACTORS AND ASSOCIATED PLANTS↗
DOE OSTI2023

Three-dimensional nanoscale reduced-angle ptycho-tomographic imaging with deep learning (RAPID)

X-ray ptychographic tomography is a nondestructive method for three dimensional (3D) imaging with nanometer-sized resolvable features. The size of the volume that can be imaged is almost arbitrary, limited only by the penetration depth and the available scanning time. Here we present a method that rapidly accelerates the imaging operation over a given volume through acquiring a limited set of data via large angular reduction and compensating for the resulting ill-posedness through deeply learned priors. The proposed 3D reconstruction method “RAPID” relies initially on a subset of the object measured with the nominal number of required illumination angles and treats the reconstructions from the conventional two-step approach as ground truth. It is then trained to reproduce equal fidelity from much fewer angles. After training, it performs with similar fidelity on the hitherto unexamined portions of the object, previously not shown during training, with a limited set of acquisitions. In our experimental demonstration, the nominal number of angles was 349 and the reduced number of angles was 21, resulting in a x140 aggregate speedup over a volume of 4.48 x 93.18 x 3.92 μm 3 and with (14nm) 3 feature size, i.e. ~ 10 8 voxels. RAPID’s key distinguishing feature over earlier attempts is the incorporation of atrous spatial pyramid pooling modules into the deep neural network framework in an anisotropic way. We found that adjusting the atrous rate improves reconstruction fidelity because it expands the convolutional kernels’ range to match the physics of multi-slice ptychography without significantly increasing the number of parameters.

47 OTHER INSTRUMENTATION↗
DOE OSTI2023

Denoising diffusion models with geometry adaptation for high fidelity calorimeter simulation

Simulation is crucial for all aspects of collider data analysis, but the available computing budget in the High Luminosity LHC era will be severely constrained. Generative machine learning models may act as surrogates to replace physics-based full simulation of particle detectors, and diffusion models have recently emerged as the state of the art for other generative tasks. Here, we introduce CaloDiffusion, a denoising diffusion model trained on the public CaloChallenge datasets to generate calorimeter showers. Our algorithm employs 3D cylindrical convolutions, which take advantage of symmetries of the underlying data representation. To handle irregular detector geometries, we augment the diffusion model with a new geometry latent mapping (GLaM) layer to learn forward and reverse transformations to a regular geometry that is suitable for cylindrical convolutions. The showers generated by our approach are nearly indistinguishable from the full simulation, as measured by several different metrics.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗
DOE OSTI2024

Local primordial non-Gaussianity from the large-scale clustering of photometric DESI luminous red galaxies

ABSTRACT We use angular clustering of luminous red galaxies from the Dark Energy Spectroscopic Instrument (DESI) imaging surveys to constrain the local primordial non-Gaussianity parameter fNL. Our sample comprises over 12 million targets, covering 14 000 deg2 of the sky, with redshifts in the range 0.2 < z < 1.35. We identify Galactic extinction, survey depth, and astronomical seeing as the primary sources of systematic error, and employ linear regression and artificial neural networks to alleviate non-cosmological excess clustering on large scales. Our methods are tested against simulations with and without fNL and systematics, showing superior performance of the neural network treatment. The neural network with a set of nine imaging property maps passes our systematic null test criteria, and is chosen as the fiducial treatment. Assuming the universality relation, we find $f_{\rm NL} = 34^{+24(+50)}_{-44(-73)}$ at 68 per cent (95 per cent) confidence. We apply a series of robustness tests (e.g. cuts on imaging, declination, or scales used) that show consistency in the obtained constraints. We study how the regression method biases the measured angular power spectrum and degrades the fNL constraining power. The use of the nine maps more than doubles the uncertainty compared to using only the three primary maps in the regression. Our results thus motivate the development of more efficient methods that avoid overcorrection, protect large-scale clustering information, and preserve constraining power. Additionally, our results encourage further studies of fNL with DESI spectroscopic samples, where the inclusion of 3D clustering modes should help separate imaging systematics and lessen the degradation in the fNL uncertainty.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2023

Universal Polarization Transformations: Spatial Programming of Polarization Scattering Matrices Using a Deep Learning‐Designed Diffractive Polarization Transformer

Abstract Controlled synthesis of optical fields having nonuniform polarization distributions presents a challenging task. Here, a universal polarization transformer is demonstrated that can synthesize a large set of arbitrarily‐selected, complex‐valued polarization scattering matrices between the polarization states at different positions within its input and output field‐of‐views (FOVs). This framework comprises 2D arrays of linear polarizers positioned between isotropic diffractive layers, each containing tens of thousands of diffractive features with optimizable transmission coefficients. After its deep learning‐based training, this diffractive polarization transformer can successfully implement N i N o = 10 000 different spatially‐encoded polarization scattering matrices with negligible error, where N i and N o represent the number of pixels in the input and output FOVs, respectively. This universal polarization transformation framework is experimentally validated in the terahertz spectrum by fabricating wire‐grid polarizers and integrating them with 3D‐printed diffractive layers to form a physical polarization transformer. Through this set‐up, an all‐optical polarization permutation operation of spatially‐varying polarization fields is demonstrated, and distinct spatially‐encoded polarization scattering matrices are simultaneously implemented between the input and output FOVs of a compact diffractive processor. This framework opens up new avenues for developing novel devices for universal polarization control and may find applications in, e.g., remote sensing, medical imaging, security, material inspection, and machine vision.

Optical neural networks↗
DOE OSTI2024

3D-Reconstruction of Tau Neutrinos in LArTPC Detectors

The Deep Underground Neutrino Experiment (DUNE) is a next-generation neutrino experiment currently under construction. DUNE will consist of two high-resolution neutrino interaction imaging detectors exposed to the world’s most intense neutrino beam, with the Near Detector at Fermilab and the Far Detector 1,300 km away in the Sanford Underground Research Facility in South Dakota, US. The high statistics and excellent resolution capabilities of DUNE's $^{40}$Ar detector will allow us to make precision studies of oscillation parameters capable of searching for CP violation in the lepton sector, testing interaction models, and studying phenomena that have until now, seemed too complex to measure, like $\nu_\tau$ detection and therefore, providing the completion of the 3-flavor neutrino paradigm. Knowledge of the $\nu_\tau$ detection can impact a broad spectrum of open questions. These include searching for non-standard neutrino interactions, constraining the unitarity of the PMNS matrix, searching for sterile neutrinos, and studying neutrino interactions. In the case of LArTPC data, the detector hits can be considered nodes in a graph, and the edges represent the spatial and temporal relationships between them. By using graph neural networks, it is possible to exploit these relationships and improve the accuracy of particle identification and reconstruction. During my presentation and specifically for tau neutrino reconstruction, I will show the effectiveness and reliability of our in-house developed graph neural network (GNN), NuGraph. This GNN classifies detector hits based on the particle type responsible for their production, assuring that the system accurately identifies and categorizes information based on its unique characteristics.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2020

Enabling Highly Efficient Capsule Networks Processing Through A PIM-Based Architecture Design

In recent years, the CNNs have achieved great successes in the image processing tasks, e.g., image recognition and object detection. Unfortunately, traditional CNN's classication is found to be easily misled by increasingly complex image features due to the usage of pooling operations, hence unable to preserve accurate position and pose information of the objects. To address this challenge, a novel neural network structure called Capsule Network has been proposed, which introduces equivariance through capsules to signicantly enhance the learning ability for image segmentation and object detection. Due to its requirement of performing a high volume of matrix operations, CapsNets have been generally accelerated on modern GPU platforms that provide highly optimized software library for common deep learning tasks. However, based on our performance characterization on modern GPUs, CapsNets exhibit low effciency due to the special program and execution features of their routing procedure, including massive unshareable intermediate variables and intensive syn- chronizations, which are very dicult to optimize at software level. To address these challenges, we propose a hybrid computing architecture design named PIM-CapsNet. It preserves GPU's on-chip computing capability for accelerating CNN types of layers in CapsNet, while pipelining with an off-chip in-memory acceleration solution that effectively tackles routing procedure's ineffciency by leveraging the processing-in-memory capability of today's 3D stacked memory. Using routing procedure's inherent parallellization feature, our design enables hierarchical improvements on CapsNet inference effciency through minimizing data movement and maximizing parallel processing in memory. Evaluation results demonstrate that our proposed design can achieve substantial improvement on both performance and energy savings for CapsNet inference, with almost zero accuracy loss. The results also suggest good performance scalability in optimizing the routing procedure with increasing network size.

Zhang, Xingyao↗