Kary Myers

Taming massive distributed datasets: data sampling using bitmap indices

With growing computational capabilities of parallel machines, scientific simulations are being performed at finer spatial and temporal scales, leading to a data explosion. The growing sizes are making it extremely hard to store, manage, disseminate, analyze, and visualize these datasets, especially as neither the memory capacity of parallel machines, memory access speeds, nor disk bandwidths are increasing at the same rate as the computing power. Sampling can be an effective technique to address the above challenges, but it is extremely important to ensure that dataset characteristics are preserved, and the loss of accuracy is within acceptable levels. In this paper, we address the data explosion problems by developing a novel sampling approach, and implementing it in a flexible system that supports server-side sampling and data subsetting. We observe that to allow subsetting over scientific datasets, data repositories are likely to use an indexing technique. Among these techniques, we see that bitmap indexing can not only effectively support subsetting over scientific datasets, but can also help create samples that preserve both value and spatial distributions over scientific datasets. We have developed algorithms for using bitmap indices to sample datasets. We have also shown how only a small amount of additional metadata stored with bitvectors can help assess loss of accuracy with a particular subsampling level. Some of the other properties of this novel approach include: 1) sampling can be flexibly applied to a subset of the original dataset, which may be specified using a value-based and/or a dimension-based subsetting predicate, and 2) no data reorganization is needed, once bitmap indices have been generated. We have extensively evaluated our method with different types of datasets and applications, and demonstrated the effectiveness of our approach.

ADR visualization: A generalized framework for ranking large-scale scientific data using Analysis-Driven Refinement

Prioritization of data is necessary for managing large-scale scientific data, as the scale of the data implies that there are only enough resources available to process a limited subset of the data. For example, data prioritization is used during in situ triage to scale with bandwidth bottlenecks, and used during focus+context visualization to save time during analysis by guiding the user to important information. In this paper, we present ADR visualization, a generalized analysis framework for ranking large-scale data using Analysis-Driven Refinement (ADR), which is inspired by Adaptive Mesh Refinement (AMR). A large-scale data set is partitioned in space, time, and variable, using user-defined importance measurements for prioritization. This process creates a prioritization tree over the data set. Using this tree, selection methods can generate sparse data products for analysis, such as focus+context visualizations or sparse data sets.

Effective and efficient data sampling using bitmap indices

With growing computational capabilities of parallel machines, scientific simulations are being performed at finer spatial and temporal scales, leading to a data explosion. The growing sizes are making it extremely hard to store, manage, disseminate, analyze, and visualize these datasets, especially as neither the memory capacity of parallel machines, memory access speeds, nor disk bandwidths are increasing at the same rate as the computing power. Sampling can be an effective technique to address the above challenges, but it is extremely important to ensure that dataset characteristics are preserved, and the loss of accuracy is within acceptable levels. In this paper, we address the data explosion problems by developing a novel sampling approach, and implementing it in a flexible system that supports server-side sampling and data subsetting. We observe that to allow subsetting over scientific datasets, data repositories are likely to use an indexing technique. Among these techniques, we see that bitmap indexing can not only effectively support subsetting over scientific datasets, but can also help create samples that preserve both value and spatial distributions over scientific datasets. We have developed algorithms for using bitmap indices to sample datasets. We have also shown how only a small amount of additional metadata stored with bitvectors can help assess loss of accuracy with a particular subsampling level. Some of the other properties of this novel approach include: (1) sampling can be flexibly applied to a subset of the original dataset, which may be specified using a value-based and/or a dimension-based subsetting predicate, and (2) no data reorganization is needed, once bitmap indices have been generated. We have extensively evaluated our method with different types of datasets and applications, and demonstrated the effectiveness of our approach.

Partitioning a Large Simulation as It Runs

As computer simulations continue to grow in size and complexity, they present a particularly challenging class of big data problems. Many application areas are moving toward exascale computing systems, systems that perform 1018 FLOPS (FLoating-point Operations Per Second)—a billion billion calculations per second. Simulations at this scale can generate output that exceeds both the storage capacity and the bandwidth available for transfer to storage, making post-processing and analysis challenging. One approach is to embed some analyses in the simulation while the simulation is running—a strategy often called in situ analysis—to reduce the need for transfer to storage. Another strategy is to save only a reduced set of time steps rather than the full simulation. Typically the selected time steps are evenly spaced, where the spacing can be defined by the budget for storage and transfer. This article combines these two ideas to introduce an online in situ method for identifying a reduced set of time steps of the simulation to save. Our approach significantly reduces the data transfer and storage requirements, and it provides improved fidelity to the simulation to facilitate post-processing and reconstruction. We illustrate the method using a computer simulation that supported NASAs 2009 Lunar Crater Observation and Sensing Satellite mission.

Sparse classification of rf transients using chirplets and learned dictionaries

We assess the performance of a sparse classification approach for radiofrequency (RF) transient signals using dictionaries adapted to the data. We explore two approaches: pursuit-type decompositions over analytical, over-complete dictionaries, and dictionaries learned directly from data. Pursuit-type decompositions over analytical, over-complete dictionaries yield sparse representations by design and can work well for target signals in the same function class as the dictionary atoms. Discriminative dictionaries learned directly from data do not rely on analytical constraints or additional knowledge about the signal characteristics, and provide sparse representations that can perform well when used with a statistical classifier. We present classification results for learned dictionaries on simulated test data, and discuss robustness compared to conventional Fourier methods. We draw from techniques of adaptive feature extraction, statistical machine learning, and image processing.

Classification of transient signals using sparse representations over adaptive dictionaries

Automatic classification of broadband transient radio frequency (RF) signals is of particular interest in persistent surveillance applications. Because such transients are often acquired in noisy, cluttered environments, and are characterized by complex or unknown analytical models, feature extraction and classification can be difficult. We propose a fast, adaptive classification approach based on non-analytical dictionaries learned from data. Conventional representations using fixed (or analytical) orthogonal dictionaries, e.g., Short Time Fourier and Wavelet Transforms, can be suboptimal for classification of transients, as they provide a rigid tiling of the time-frequency space, and are not specifically designed for a particular signal class. They do not usually lead to sparse decompositions, and require separate feature selection algorithms, creating additional computational overhead. Pursuit-type decompositions over analytical, redundant dictionaries yield sparse representations by design, and work well for target signals in the same function class as the dictionary atoms. The pursuit search however has a high computational cost, and the method can perform poorly in the presence of realistic noise and clutter. Our approach builds on the image analysis work of Mairal et al. (2008) to learn a discriminative dictionary for RF transients directly from data without relying on analytical constraints or additional knowledge about the signal characteristics. We then use a pursuit search over this dictionary to generate sparse classification features. We demonstrate that our learned dictionary is robust to unexpected changes in background content and noise levels. The target classification decision is obtained in almost real-time via a parallel, vectorized implementation.

Radio frequency (RF) transient classification using sparse representations over learned dictionaries

Automatic classification of transitory or pulsed radio frequency (RF) signals is of particular interest in persistent surveillance and remote sensing applications. Such transients are often acquired in noisy, cluttered environments, and may be characterized by complex or unknown analytical models, making feature extraction and classification difficult. We propose a fast, adaptive classification approach based on non-analytical dictionaries learned from data. We compare two dictionary learning methods from the image analysis literature, the K-SVD algorithm and Hebbian learning, and extend them for use with RF data. Both methods allow us to learn discriminative RF dictionaries directly from data without relying on analytical constraints or additional knowledge about the expected signal characteristics. We then use a pursuit search over the learned dictionaries to generate sparse classification features in order to identify time windows that contain a target pulse. In this paper we compare the two dictionary learning methods and discuss how their performance changes as a function of dictionary training parameters. We demonstrate that learned dictionary techniques are suitable for pulsed RF analysis and present results with varying background clutter and noise levels.

Effects of background suppression of gamma counts on signal estimation

Gamma detectors at border crossings are intended to detect illicit nuclear material. One of their performance challenges is the fact that vehicles suppress the natural background and, thus, potentially reduce probability of detection of threat items. Here we test several methods to adjust the detection to background suppression in the context of signal estimation. We show that, for the small-to-moderate suppression magnitudes, suppression adjustment leads to higher detection probability. However, for signals triggering alarm without an adjustment, adjustment does not improve estimation of the signal location, only moderately improves estimation of the signal magnitude, and does not improve estimation of the signal width.

Change Detection Using Wavelets in Solution Monitoring Data for Nuclear Safeguards

Wavelet analysis is known to be a good option for change detection in many contexts. Detecting changes in solution volumes that are measured with both additive and relative error is an important aspect of safeguards for facilities that process special nuclear material. This paper qualitatively compares wavelet-based change detection to a lag-one differencing option using realistic simulated solution volume data for which the true change points are known. We then show quantitatively that Haar wavelet-based change detection is effective for finding the approximate location of each change point, and that a simple piecewise linear optimization step is effective to refine the initial wavelet-based change point estimate.

Capturing dynamics on multiple time scales: a multilevel fusion approach for cluttered electromagnetic data

Many problems in electromagnetic signal analysis exhibit dynamics on a wide range of time scales. Further, these dynamics may involve both continuous source generation processes and discrete source mode dynamics. These rich temporal characteristics can present challenges for standard modeling approaches, particularly in the presence of nonstationary noise and clutter sources. Here we demonstrate a hybrid algorithm designed to capture the dynamic behavior at all relevant time scales while remaining robust to clutter and noise at each time scale. We draw from techniques of adaptive feature extraction, statistical machine learning, and discrete process modeling to construct our hybrid algorithm. We describe our approach and present results applying our hybrid algorithm to a simulated dataset based on an example radio beacon identification problem: civilian air traffic control. This application illustrates the multi-scale complexity of the problems we wish to address. We consider a multi-mode air traffic control radar emitter operating against a cluttered background of competing radars and continuous-wave communications signals (radios, TV broadcasts). Our goals are to find a compact representation of the radio frequency measurements, identify which pulses were emitted by the target source, and determine the mode of the source.