Search NASA⌕ Search

SEARCH · Search NASA

Results for “Labeled Datasets”

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 541 records · Page 30

MABAL: a Novel Deep-Learning Architecture for Machine-Assisted Bone Age Labeling

Bone age assessment (BAA) is a commonly performed diagnostic study in pediatric radiology to assess skeletal maturity. The most commonly utilized method for assessment of BAA is the Greulich and Pyle method (Pediatr Radiol 46.9:1269–1274, 2016; Arch Dis Child 81.2:172–173, 1999) atlas. The evaluation of BAA can be a tedious and time-consuming process for the radiologist. As such, several computer-assisted detection/diagnosis (CAD) methods have been proposed for automation of BAA. Classical CAD tools have traditionally relied on hard-coded algorithmic features for BAA which suffer from a variety of drawbacks. Recently, the advent and proliferation of convolutional neural networks (CNNs) has shown promise in a variety of medical imaging applications. There have been at least two published applications of using deep learning for evaluation of bone age (Med Image Anal 36:41–51, 2017; JDI 1–5, 2017). However, current implementations are limited by a combination of both architecture design and relatively small datasets. The purpose of this study is to demonstrate the benefits of a customized neural network algorithm carefully calibrated to the evaluation of bone age utilizing a relatively large institutional dataset. In doing so, this study will aim to show that advanced architectures can be successfully trained from scratch in the medical imaging domain and can generate results that outperform any existing proposed algorithm. The training data consisted of 10,289 images of different skeletal age examinations, 8909 from the hospital Picture Archiving and Communication System at our institution and 1383 from the public Digital Hand Atlas Database. The data was separated into four cohorts, one each for male and female children above the age of 8, and one each for male and female children below the age of 10. The testing set consisted of 20 radiographs of each 1-year-age cohort from 0 to 1 years to 14–15+ years, half male and half female. The testing set included left-hand radiographs done for bone age assessment, trauma evaluation without significant findings, and skeletal surveys. A 14 hidden layer-customized neural network was designed for this study. The network included several state of the art techniques including residual-style connections, inception layers, and spatial transformer layers. Data augmentation was applied to the network inputs to prevent overfitting. A linear regression output was utilized. Mean square error was used as the network loss function and mean absolute error (MAE) was utilized as the primary performance metric. MAE accuracies on the validation and test sets for young females were 0.654 and 0.561 respectively. For older females, validation and test accuracies were 0.662 and 0.497 respectively. For young males, validation and test accuracies were 0.649 and 0.585 respectively. Finally, for older males, validation and test set accuracies were 0.581 and 0.501 respectively. The female cohorts were trained for 900 epochs each and the male cohorts were trained for 600 epochs. An eightfold cross-validation set was employed for hyperparameter tuning. Test error was obtained after training on a full data set with the selected hyperparameters. Using our proposed customized neural network architecture on our large available data, we achieved an aggregate validation and test set mean absolute errors of 0.637 and 0.536 respectively. To date, this is the best published performance on utilizing deep learning for bone age assessment. Our results support our initial hypothesis that customized, purpose-built neural networks provide improved performance over networks derived from pre-trained imaging data sets. We build on that initial work by showing that the addition of state-of-the-art techniques such as residual connections and inception architecture further improves prediction accuracy. This is important because the current assumption for use of residual and/or inception architectures is that a large pre-trained network is required for successful implementation given the relatively small datasets in medical imaging. Instead we show that a small, customized architecture incorporating advanced CNN strategies can indeed be trained from scratch, yielding significant improvements in algorithm accuracy. It should be noted that for all four cohorts, testing error outperformed validation error. One reason for this is that our ground truth for our test set was obtained by averaging two pediatric radiologist reads compared to our training data for which only a single read was used. This suggests that despite relatively noisy training data, the algorithm could successfully model the variation between observers and generate estimates that are close to the expected ground truth.

62 RADIOLOGY AND NUCLEAR MEDICINE↗

Improving Sim-to-Real Transfer in Vision-Based Robot Navigation Via Instance-Level GAN-Based Data Augmentation

Achieving robust vision-based robotic tasks requires large amounts of data, which are often difficult to obtain in real-world scenarios. Simulators and synthetic data offer a cost-effective alternative, but the visual gap between simulation and reality hinders the performance of models when deployed in real-world environments. In this paper, we present a data augmentation pipeline that integrates a foundation model (Segment Anything Model) with an unsupervised image-to-image translation model (CycleGAN) for instance-level domain transfer from simulation to reality. This pipeline enables the generation of realistic labeled data from synthetic images for training supervised machine learning models in vision-based navigation tasks. We evaluate our approach on real-world data for ego-vehicle pose estimation, a critical autonomous navigation task involving the prediction of cross-track position and heading angle relative to road center line markings. The results of our tests show that our GAN-based data augmentation pipeline significantly outperforms models trained solely on simulation data or on data processed with standard image augmentation methods for sim-to-real transfer, enhancing model robustness and generalizability in real-world scenarios. Our method provides a scalable and flexible data augmentation tool for leveraging large synthetic datasets to enhance vision-based robotic navigation tasks.

artificial intelligence↗

SU-E-J-212: Identifying Bones From MRI: A Dictionary Learnign and Sparse Regression Approach

Purpose: To develop an efficient and robust scheme to identify bony anatomy based on MRI-only simulation images. Methods: MRI offers important soft tissue contrast and functional information, yet its lack of correlation to electron-density has placed it as an auxiliary modality to CT in radiotherapy simulation and adaptation. An effective scheme to identify bony anatomy is an important first step towards MR-only simulation/treatment paradigm and would satisfy most practical purposes. We utilize a UTE acquisition sequence to achieve visibility of the bone. By contrast to manual + bulk or registration-to identify bones, we propose a novel learning-based approach for improved robustness to MR artefacts and environmental changes. Specifically, local information is encoded with MR image patch, and the corresponding label is extracted (during training) from simulation CT aligned to the UTE. Within each class (bone vs. nonbone), an overcomplete dictionary is learned so that typical patches within the proper class can be represented as a sparse combination of the dictionary entries. For testing, an acquired UTE-MRI is divided to patches using a sliding scheme, where each patch is sparsely regressed against both bone and nonbone dictionaries, and subsequently claimed to be associated with the class with the smaller residual. Results: The proposed method has been applied to the pilot site of brain imaging and it has showed general good performance, with dice similarity coefficient of greater than 0.9 in a crossvalidation study using 4 datasets. Importantly, it is robust towards consistent foreign objects (e.g., headset) and the artefacts relates to Gibbs and field heterogeneity. Conclusion: A learning perspective has been developed for inferring bone structures based on UTE MRI. The imaging setting is subject to minimal motion effects and the post-processing is efficient. The improved efficiency and robustness enables a first translation to MR-only routine. The scheme generalizes to multiple tissue classes.

60 APPLIED LIFE SCIENCES↗

Deep Analysis Net with Causal Embedding for Coal-fired Power Plant Fault Detection and Diagnosis (DANCE4CFDD)

Fault detection and diagnosis is critical to power plant operation to ensure attaining high reliability while reducing operation cost. As more renewable power is introduced to the power grid, traditional fossil power plants take on the extra burden of excessive load cycling to compensate the generation variability from renewable power. Such load cycling will pose more reliability challenges to power plant operation. There are a number of challenges faced by today’s asset health management system in coal- fired (or gas) power plants: 1) high-dimensional nonlinear interaction among multiple time series measurements; 2) high measurement variance induced by operational conditions/modes; 3) variation among asset types and plant configurations; and 4) a small number of faulty events to learn from. To cope with these challenges, today’s fielded asset health management systems rely heavily on manual efforts from domain experts and hand-crafted features or rules based on domain knowledge. Despite its role in plant reliability, such a practice is costly and hinders its scalability and sustainability, particularly when a plant undergoes modifications. The objective of this project is to develop a novel end-to-end AI learning system that is trainable (i.e., the AI representation of a complex system behavior can be directly learned from properly labeled data) for accurate fault detection and root cause analysis. The ability to create a fault detection model directly from time series could alleviate the efforts associated with today’s asset management solution development. In the course of this project, we have achieved the following: Created an AI model development environment incorporating state-of-the-art neural network architectures for rapid model development and evaluation; Developed novel learning strategies for training of fault detection model; Developed special-purpose neural network architecture embedded with variable association graph aiming for better interpretability; Developed a learning strategy to leverage a small number of faulty events for enhanced fault detection capability; Conducted detailed experimental study based on public benchmark datasets and demonstrated the effectiveness of the proposed solution; and Validated the developed system with data from both a coal-fired plant boiler dynamic simulation model and real-world coal-fired power plant covering multiple asset and fault types. Overall, the project attained a technology readiness level of TRL 5 from TRL 2 at the beginning of the project.

20 FOSSIL-FUELED POWER PLANTS↗

From genomes to metabolomes: Understanding mechanisms of symbiosis and cell-cell signaling using the archaeal system Ignicoccus-Nanoarchaeum

The main objective of this project was to use symbiotic Nanoarchaeaota, a group of thermophilic Archaea that are obligate symbionts/parasites on other Archaea, to develop an integrated multi-omic approach to study inter-species interactions as well as to understand fundamental mechanism that enable such relationships. As part of this grant we have achieved a number of important milestone on both technical and scientific levels. On the technical side, we developed immunofluorescence labeling and tracking methods to follow Nanoarchaeota in cultures and in environmental samples, we applied such methods in conjunction with flow cytometry to quantify and isolate uncultured representatives from the environment and characterized them by single cell genomics. On the proteomics side, we developed a more efficient and sensitive method to recover and semi-quantitatively measure membrane proteins, while achieving high total cellular proteome coverage (70-80% of the predicted proteome). Metabolomic analyses used complementary NMR and LC/GC mass spectrometry and led to the identification of novel lipids in these organisms as well as quantification of some of the major metabolites. Importantly, using several informatics approaches we were also able to integrate the transcriptomic, proteomic and metabolomic datasets, revealing aspects of the interspecies interaction that were not evident in the single omic analyses (manuscript in review). On the science side we determined that N. equitans and I. hospitalis are metabolically coupled and that N. equitans is strictly dependent on its host both for metabolic precursors and energetic needs. The actual mechanism by which small molecules move across the cell membrane remains unknown. The Ignicoccus host responds to the metabolic and energetic burned by upregulating of key primary metabolism steps and ATP synthesis. The two species have co-evolved, aspect that we determined by comparative genomics with other species of Ignicoccus (manuscript in preparation) and by characterizing other similar Nanoarchaeota systems. Using a single cell genomics approach we characterized the first terrestrial geothermal Nanoarchaeota system, from Yellowstone National Park. That nanoarchaeon uses a different host, a species of Sulfolobales, and comparative genomics with N. equitans-Ignicoccus allowed us to come up with an evolutionary model for the evolution of this group of organisms across marine and terrestrial ecosystems. Based on metabolic inferences we were also able to isolate in culture the first such terrestrial nanoarchaeal system, also from Yellowstone, which involves a species of Acidilobus. The novel nanoarchaeal system was characterized using proteomics and it helped us better understand the metabolic capabilities of these organisms as well as how co-evolution shapes the genomes of interacting species. It was also one of the very few cases in which prior genomic data was used to successfully design an approach to culture an organism, which remains the gold standard in microbiology research. As a better understanding of interspecies interaction requires multiple model systems, we have pursued identification and genomic characterization or isolation of additional nanoarchaeal systems from geographically and geochemically distinct environments. Two additional nanoarchaeal systems are presently being characterized from hot springs in Yellowstone and Iceland and will be the subject to future publications.

59 BASIC BIOLOGICAL SCIENCES↗

Machine Learning Automated Analysis of Enormous Synchrotron X-ray Diffraction Datasets

X-ray diffraction (XRD) data analysis can be a time-consuming and laborious task. Deep neural network (DNN) based models trained with synthetic XRD patterns have been proven to be a highly efficient, accurate, and automated method for analyzing common XRD data collected from solid samples in ambient environments. However, it remains unclear whether synthetic XRD-based models can be effective in solving micro(μ)-XRD mapping data for in situ experiments involving liquid phases, which always have lower quality and significant artifacts. In this study, we collected μ-XRD mapping data from a LaCl 3 -calcite hydrothermal fluid system and trained two categories of models to analyze the experimental XRD patterns. Here, the models trained solely with synthetic XRD patterns showed low accuracy (as low as 64%) when solving experimental μ-XRD mapping data. However, the accuracy of the DNN models significantly improved (90% or above) when we trained them with a data set containing both synthetic and a small number of labeled experimental μ-XRD patterns. This study highlights the importance of labeled experimental patterns in training DNN models to solve μ-XRD mapping data from in situ experiments involving liquid phases.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

A High-Throughput Solver for Marginalized Graph Kernels on GPU

Here, we present the design and optimization of a solver for efficient and high-throughput computation of the marginalized graph kernel on General Purpose GPUs. The graph kernel is computed using the conjugate gradient method to solve a generalized Laplacian of the tensor product between a pair of graphs. To cope with the large gap between the instruction throughput and the memory bandwidth of the GPUs, our solver forms the graph tensor product on-the-fly without storing it in memory. This is achieved by using threads in a warp cooperatively to stream the adjacency and edge label matrices of individual graphs by small square matrix blocks called tiles, which are then staged in registers and the shared memory for later reuse. Warps across a thread block can further share tiles via the shared memory to increase data reuse. We exploit the sparsity of the graphs hierarchically by storing only non-empty tiles using a coordinate format and nonzero elements within each tile using bitmaps. We propose a new partition-based reordering algorithm for aggregating nonzero elements of the graphs into fewer but denser tiles to further exploit sparsity. We carry out extensive theoretical analyses on the graph tensor product primitives for tiles of various density and evaluate their performance on synthetic and real-world datasets. Our solver delivers three to four orders of magnitude speedup over existing CPU-based solvers such as GraKeL and GraphKernels. The capability of the solver enables kernel-based learning tasks at unprecedented scales.

97 MATHEMATICS AND COMPUTING↗

Active Learning with Irrelevant Examples

Active learning algorithms attempt to accelerate the learning process by requesting labels for the most informative items first. In real-world problems, however, there may exist unlabeled items that are irrelevant to the user's classification goals. Queries about these points slow down learning because they provide no information about the problem of interest. We have observed that when irrelevant items are present, active learning can perform worse than random selection, requiring more time (queries) to achieve the same level of accuracy. Therefore, we propose a novel approach, Relevance Bias, in which the active learner combines its default selection heuristic with the output of a simultaneously trained relevance classifier to favor items that are likely to be both informative and relevant. In our experiments on a real-world problem and two benchmark datasets, the Relevance Bias approach significantly improved the learning rate of three different active learning approaches.

machine learning↗

Convolutional networks for vehicle track segmentation

Existing methods to detect vehicle tracks in coherent change detection images, a product of combining two synthetic aperture radar images taken at different times of the same scene, rely on simple, fast models to label track pixels. These models, however, are unable to capture natural track features such as continuity and parallelism. More powerful, but computationally expensive models can be used in offline settings. We present an approach that uses dilated convolutional networks consisting of a series of 3-by-3 convolutions to segment vehicle tracks. The design of our networks considers the fact that remote sensing applications tend to operate in low power and have limited training data. As a result, we aim for small, efficient networks that can be trained end-to-end to learn natural track features entirely from limited training data. We demonstrate that our 6-layer network, trained on just 90 images, is computationally efficient and improves the F-score on a standard dataset to 0.992, up from 0.959 obtained by the current state-of-the-art method.

47 OTHER INSTRUMENTATION↗

The Ground Flash Fraction Retrieval Algorithm Employing Differential Evolution: Simulations and Applications

The ability to estimate the fraction of ground flashes in a set of flashes observed by a satellite lightning imager, such as the future GOES-R Geostationary Lightning Mapper (GLM), would likely improve operational and scientific applications (e.g., severe weather warnings, lightning nitrogen oxides studies, and global electric circuit analyses). A Bayesian inversion method, called the Ground Flash Fraction Retrieval Algorithm (GoFFRA), was recently developed for estimating the ground flash fraction. The method uses a constrained mixed exponential distribution model to describe a particular lightning optical measurement called the Maximum Group Area (MGA). To obtain the optimum model parameters (one of which is the desired ground flash fraction), a scalar function must be minimized. This minimization is difficult because of two problems: (1) Label Switching (LS), and (2) Parameter Identity Theft (PIT). The LS problem is well known in the literature on mixed exponential distributions, and the PIT problem was discovered in this study. Each problem occurs when one allows the numerical minimizer to freely roam through the parameter search space; this allows certain solution parameters to interchange roles which leads to fundamental ambiguities, and solution error. A major accomplishment of this study is that we have employed a state-of-the-art genetic-based global optimization algorithm called Differential Evolution (DE) that constrains the parameter search in such a way as to remove both the LS and PIT problems. To test the performance of the GoFFRA when DE is employed, we applied it to analyze simulated MGA datasets that we generated from known mixed exponential distributions. Moreover, we evaluated the GoFFRA/DE method by applying it to analyze actual MGAs derived from low-Earth orbiting lightning imaging sensor data; the actual MGA data were classified as either ground or cloud flash MGAs using National Lightning Detection Network[TM] (NLDN) data. Solution error plots are provided for both the simulations and actual data analyses.

Koshak, William↗

Semi-supervised Bayesian Low-shot Learning

Deep neural networks (NNs) typically outperform traditional machine learning (ML) approaches for complicated, non-linear tasks. It is expected that deep learning (DL) should offer superior performance for the important non-proliferation task of predicting explosive device configuration based upon observed optical signature, a task which human experts struggle with. However, supervised machine learning is difficult to apply in this mission space because most recorded signatures are not associated with the corresponding device description, or “truth labels.” This is challenging for NNs, which traditionally require many samples for strong performance. Semi-supervised learning (SSL), low-shot learning (LSL), and uncertainty quantification (UQ) for NNs are emerging approaches that could bridge the mission gaps of few labels and rare samples of importance. NN explainability techniques are important in gaining insight into the inferential feature importance of such a complex model. In this work, SSL, LSL, and UQ are merged into a single framework, a significant technical hurdle not previously demonstrated. Exponential Average Adversarial Training (EAAT) and Pairwise Neural Networks (PNNs) are chosen as the SSL and LSL methods of choice. Permutation feature importance (PFI) for functional data is used to provide explainability via the Variable importance Explainable Elastic Shape Analysis (VEESA) pipeline. A variety of uncertainty quantification approaches are explored: Bayesian Neural Networks (BNNs), ensemble methods, concrete dropout, and evidential deep learning. Two final approaches, one utilizing ensemble methods and one utilizing evidential learning, are constructed and compared using a well-quantified synthetic 2D dataset along with the DIRSIG Megascene.

97 MATHEMATICS AND COMPUTING↗

Community Geothermal: Soil Conductivity, Borehole Design, Energy Models, and Load Data for a Residential System Development - Hinesburg, VT

This dataset contains materials from the Coalition for Community-Supported Affordable Geothermal Energy Systems (C2SAGES) project, which evaluated the techno-economic feasibility of a community geothermal system for a residential development in Hinesburg, VT. The dataset includes detailed soil conductivity test reports, energy models, borehole design reports, hourly energy loads for heating, cooling, and hot water, and design layouts. EnergyPlus was used to model building energy loads, and Modelica software was applied for geothermal loop sizing based on these loads and soil conductivity results. Python scripts for network design further refined the models. Key files include PDF reports on borehole design (with projections for 1-year, 15-year, and 30-year systems), soil conductivity test results, EnergyPlus modeling outputs, and 2D/3D design drawings in PDF, DWG, and DXF formats. Python notebooks for network design and OnePipe model files are also provided, with Modelica required for viewing certain files. Outputs and modeling data are in various formats including CSV, JPG, HTML, and IDF, with units and data clearly labeled to support understanding of system design and performance for the proposed geothermal solution.

15 GEOTHERMAL ENERGY↗

Supervised Gamma Process Poisson Factorization

This thesis develops the supervised gamma process Poisson factorization (S- GPPF) framework, a novel supervised topic model for joint modeling of count matrices and document labels. S-GPPF is fully generative and nonparametric: document labels and count matrices are modeled under a uni ed probabilistic framework and the number of latent topics is controlled automatically via a gamma process prior. The framework provides for multi-class classification of documents using a generative max-margin classifier. Several recent data augmentation techniques are leveraged to provide for exact inference using a Gibbs sampling scheme. The first portion of this thesis reviews supervised topic modeling and several key mathematical devices used in the formulation of S-GPPF. The thesis then introduces the S-GPPF generative model and derives the conditional posterior distributions of the latent variables for posterior inference via Gibbs sampling. The S-GPPF is shown to exhibit state-of-the-art performance for joint topic modeling and document classification on a dataset of conference abstracts, beating out competing supervised topic models. The unique properties of S-GPPF along with its competitive performance make it a novel contribution to supervised topic modeling.

97 MATHEMATICS AND COMPUTING↗

Active predation, phylogenetic diversity, and global prevalence of myxobacteria in wastewater treatment plants

The operation of modern wastewater treatment plants (WWTPs) is driven by activated sludge microbiota, a complex assemblage of trophically interacting microorganisms. Microbial predation is crucial to fundamental understanding of how biological interactions drive microbiome structuring and functioning of WWTPs. However, predatory bacteria have received little attention regarding their diversity, activity, and ecological function in activated sludge, limiting the exploitation of food web interactions for wastewater microbiome engineering. Here, by using rRNA-stable isotope probing of activated sludge microbiota with 13 C-labeled prey bacteria, we uncovered diverse as-yet-uncultivated putative predatory bacteria that actively incorporated 13 C-biomass. Myxobacteria, especially Haliangium and the mle1-27 clade, were found as the dominant active predators, refreshing conventional views based on a few predatory isolates of Bdellovibrionota from WWTPs. The identified predatory bacteria showed more selective predation on prey compared with the protists dominated by ciliates, providing in situ evidence for inter-domain predation behavior divergence in activated sludge. Putative predatory bacteria were tracked over a two-year microbiome monitoring effort at a local WWTP, revealing the predominance of Myxococcota (6.5 ± 1.3%) over Bdellovibrionota (1.0 ± 0.2%) lineages. Phylogenetic analysis unveiled highly diverse myxobacteria inhabiting activated sludge and suggested a habitat filtering effect in global WWTPs. Further mining of a global activated sludge microbiome dataset revealed the prevalence of Myxococcota (5.4 ± 0.1%) species and potential impacts of myxobacterial predation on process performance. Collectively, our findings provided unique insights into the predating activity, diversity, and prevalence of Myxococcota species in activated sludge, highlighting their links with wastewater treatment processes via trophic regulation of enteric and functional bacteria.

59 BASIC BIOLOGICAL SCIENCES↗

Improving an Acoustic Vehicle Detector Using an Iterative Self-Supervision Procedure

In many non-canonical data science scenarios, obtaining, detecting, attributing, and annotating enough high-quality training data is the primary barrier to developing highly effective models. Moreover, in many problems that are not sufficiently defined or constrained, manually developing a training dataset can often overlook interesting phenomena that should be included. To this end, we have developed and demonstrated an iterative self-supervised learning procedure, whereby models are successfully trained and applied to new data to extract new training examples that are added to the corpus of training data. Successive generations of classifiers are then trained on this augmented corpus. Using low-frequency acoustic data collected by a network of infrasound sensors deployed around the High Flux Isotope Reactor and Radiochemical Engineering Development Center at Oak Ridge National Laboratory, we test the viability of our proposed approach to develop a powerful classifier with the goal of identifying vehicles from continuously streamed data and differentiating these from other sources of noise such as tools, people, airplanes, and wind. Using a small collection of exhaustively manually labeled data, we test several implementation details of the procedure and demonstrate its success regardless of the fidelity of the initial model used to seed the iterative procedure. Finally, we demonstrate the method’s ability to update a model to accommodate changes in the data-generating distribution encountered during long-term persistent data collection.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗

A Compound Data Poisoning Technique with Significant Adversarial Effects on Transformer-based Sentiment Classification Tasks

Transformer-based models have demonstrated much success in various natural language processing tasks. However, they are often vulnerable to adversarial attacks, such as data poisoning, which can intentionally fool the model into generating incorrect results. In this article, we present a novel, compound variant of a data poisoning attack on a transformer-based model that maximizes the poisoning effect while minimizing the scope of poisoning. Here we do so by combining the established data poisoning technique (label flipping) with a novel adversarial artifact selection and insertion technique aimed at minimizing detectability and the scope of the poisoning footprint. We find that by using a combination of these two techniques, we achieve a state-of-the-art attack success rate of approximately 90% while poisoning only 0.5% of the original training set, thus minimizing the scope and detectability of the poisoning action. These findings have the potential to advance the development of better data poisoning detection methods.

97 MATHEMATICS AND COMPUTING↗

Semi-Supervised Eigenbasis Novelty Detection

Recent discoveries in high-time-resolution radio astronomy data have focused attention on a new class of events. Fast transients are rare pulses of radio frequency energy lasting from microseconds to seconds that might be produced by a variety of exotic astrophysical phenomena. For example, X-ray bursts, neutron stars, and active galactic nuclei are all possible sources of short-duration, transient radio signals. It is difficult to anticipate where such signals might appear, and they are most commonly discovered through analysis of high-time- resolution data that had been collected for other purposes. Transients are often faint and difficult to detect, so improved detection algorithms can directly benefit the science yield of all such commensal monitoring. A new detection algorithm learns a low-dimensional linear manifold for describing the normal data. High reconstruction error indicates a novel signal that does not match the patterns of normal data. One unsupervised portion of the manifold model adapts its representation in response to recent data. A second supervised portion of the model is made of a basis trained in advance using labeled examples of RFI; this prevents false positives due to these events. For a linear model, an orthonormalization operation is used to combine these bases prior to the anomaly detection decision. Another novel aspect of the approach lies in combining basis vectors learned in an unsupervised, online fashion from the data stream with supervised basis vectors learned in advance from known examples of false alarms. Adaptive, data-driven detection is achieved that is also informed by existing domain knowledge about signals that may be statistically anomalous, but are not interesting and should therefore be ignored. The method was evaluated using data from the Parkes Multibeam Survey. This data set was originally collected to search for pulsars, which are astronomical sources that emit radio pulses at regular periods. However, several non-pulsar anomalies have recently been discovered in this dataset, making it a compelling test case. By explicitly filtering known false alarm patterns, the approach yields significantly better performance than current transient detection methods.

Wagstaff, Kiri L.↗

How Low Can You Go? Using Synthetic 3D Imagery to Drastically Reduce Real-World Training Data for Object Detection

Deep convolutional neural networks (DCNNs) currently provide state-of-the-art performance on image classification and object detection tasks, and there are many global security mission areas where such models could be extremely useful. Crucially, the success of these models is driven in large part by the widespread availability of high-quality open source data sets such as Image Net, Common Objects in Context (COCO), and KITTI, which contain millions of images with thousands of unique labels. However, global security relevant objects-of-interest can be difficult to obtain: relevant events are low frequency and high consequence; the content of relevant images is sensitive; and adversaries and proliferators seek to obscure their activities. For these cases where exemplar data is hard to come-by, even fine-tuning an existing model with available data can be effectively impossible. Recent work demonstrated that models can be trained using a combination of real-world and synthetic images generated from 3D representations; that such models can exceed the performance of models trained using real-world data alone; and that the generated images need not be perfectly realistic (Tremblay, et al., 2018). However, this approach still required hundreds to thousands of real-world images for training and fine tuning, which for sparse, global security-relevant datasets can be an unrealistic hurdle. In this research, we validate the performance and behavior of DCNN models as we drive the number of real-world images used for training object detection tasks down to a minimal set. We perform multiple experiments to identify the best approach to train DCNNs from an extremely small set of real-world images. In doing so, we: Develop state-of-the-art, parameterized 3D models based on real-world images and sample from their parameters to increase the variance in synthetic image training data; Use machine learning explainability techniques to highlight and correct through targeted training the biases that result from training using completely synthetic images; and Validate our results by comparing the performance of the models trained on synthetic data to one another, and to a control model created by fine-tuning an existing ImageNet-trained model with a limited number (hundreds) of real-world images.

97 MATHEMATICS AND COMPUTING↗