Search NASA⌕ Search

SEARCH · Search NASA

Results for “deep Gaussian processes”

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 127 records · Page 7

Two-photon absorption measurements of deep UV transmissible materials at 213 nm

We report on two photon absorption measurements at 213nm of deep UV transmissible media including LiF, MgF 2 , CaF 2 , BaF 2 , Sapphire (Al 2 O 3 ) and high purity grades of fused-silica (SiO 2 ). A high stability 24ps Nd:YAG laser operating at the 5th harmonic (213nm) was used to generate a high intensity, long Rayleigh length Gaussian focus inside the samples. The measurements of the Fluoride crystals and Sapphire indicate two photon absorption coefficients between 0.004 and 0.82 cm/GW. We find that different grades of fused silica performed near identically for two photon absorption, however, there are differences in linear losses associated with purity. A low two photon absorption cross section is measured for MgF 2 making it an ideal material for the propagation of high intensity deep UV lasers.

36 MATERIALS SCIENCE↗

Deep Learning of Dark Energy Spectroscopic Instrument Mock Spectra to Find Damped Lyα Systems

We have updated and applied a convolutional neural network (CNN) machine-learning model to discover and characterize damped Ly α systems (DLAs) based on Dark Energy Spectroscopic Instrument (DESI) mock spectra. We have optimized the training process and constructed a CNN model that yields a DLA classification accuracy above 99% for spectra that have signal-to-noise ratios (S/N) above 5 per pixel. The classification accuracy is the rate of correct classifications. This accuracy remains above 97% for lower S/N ≈1 spectra. This CNN model provides estimations for redshift and H i column density with standard deviations of 0.002 and 0.17 dex for spectra with S/N above 3 pixel -1 . Also, this DLA finder is able to identify overlapping DLAs and sub-DLAs. Further, the impact of different DLA catalogs on the measurement of baryon acoustic oscillations (BAO) is investigated. The cosmological fitting parameter result for BAO has less than 0.61% difference compared to analysis of the mock results with perfect knowledge of DLAs. This difference is lower than the statistical error for the first year estimated from the mock spectra: above 1.7%. We also compared the performances of the CNN and Gaussian Process (GP) models. Our improved CNN model has moderately 14% higher purity and 7% higher completeness than an older version of the GP code, for S/N > 3. Both codes provide good DLA redshift estimates, but the GP produces a better column density estimate by 24% less standard deviation. A credible DLA catalog for the DESI main survey can be provided by combining these two algorithms.

79 ASTRONOMY AND ASTROPHYSICS↗

On Transfer Learning Of Neural Networks Using Bi-Fidelity Data For Uncertainty Propagation

Due to their high degree of expressiveness, neural networks have recently been used as surrogate models for mapping inputs of an engineering system to outputs of interest. Once trained, neural networks are computationally inexpensive to evaluate and remove the need for repeated evaluations of computationally expensive models in uncertainty quantification applications. However, given the highly parameterized construction of neural networks, especially deep neural networks, accurate training often requires large amounts of simulation data that may not be available in the case of computationally expensive systems. In this paper, to alleviate this issue for uncertainty propagation, we explore the application of transfer learning techniques using training data generated from both high- and low-fidelity models. We explore two strategies for coupling these two datasets during the training procedure, namely, the standard transfer learning and the bi-fidelity-weighted learning. In the former approach, a neural network model mapping the inputs to the outputs of interest is trained based on the low-fidelity data. The high-fidelity data are then used to adapt the parameters of the upper layer(s) of the low-fidelity network, or train a simpler neural network to map the output of the low-fidelity network to that of the high-fidelity model. In the latter approach, the entire low-fidelity network parameters are updated using data generated via a Gaussian process model trained with a small high-fidelity dataset. The parameter updates are performed via a variant of stochastic gradient descent with learning rates given by the Gaussian process model. Using three numerical examples, we illustrate the utility of these bi-fidelity transfer learning methods where we focus on accuracy improvement achieved by transfer learning over standard training approaches.

97 MATHEMATICS AND COMPUTING↗

DeepCMB: Lensing Reconstruction of the CMB with Deep Neural Networks

Next-generation cosmic microwave background (CMB) experiments will have lower noise and therefore increased sensitivity, enabling improved constraints on fundamental physics parameters such as the sum of neutrino masses and the tensor-to-scalar ratio r. Achieving competitive constraints on these parameters requires high signal-to-noise extraction of the projected gravitational potential from the CMB maps. Standard methods for reconstructing the lensing potential employ the quadratic estimator (QE). However, the QE performs suboptimally at the low noise levels expected in upcoming experiments. Other methods, like maximum likelihood estimators (MLE), are under active development. In this work, we demonstrate reconstruction of the CMB lensing potential with deep convolutional neural networks (CNN) - ie, a ResUNet. The network is trained and tested on simulated data, and otherwise has no physical parametrization related to the physical processes of the CMB and gravitati onal lensing. We show that, over a wide range of angular scales, ResUNets recover the input gravitational potential with a higher signal-to-noise ratio than the QE method, reaching levels comparable to analytic approximations of MLE methods. We demonstrate that the network outputs quantifiably different lensing maps when given input CMB maps generated with different cosmologies. We also show we can use the reconstructed lensing map for cosmological parameter estimation. This application of CNN provides a few innovations at the intersection of cosmology and machine learning. First, while training and regressing on images, we predict a continuous-variable field rather than discrete classes. Second, we are able to establish uncertainty measures for the network output that are analogous to standard methods. We expect this approach to excel in capturing hard-to-model non-Gaussian astrophysical foreground and noise contributions.

79 ASTRONOMY AND ASTROPHYSICS↗

HALOGAS Observations of NGC 4559: Anomalous and Extraplanar H i and its Relation to Star Formation

We use new deep 21 cm H i observations of the moderately inclined galaxy NGC 4559 in the HALOGAS survey to investigate the properties of extraplanar gas. We use TiRiFiC to construct simulated data cubes to match the H i observations. We find that a thick-disk component of scale height ∼2 kpc, characterized by a negative vertical gradient in its rotation velocity (lag) of ∼13 ± 5 km s{sup −1} kpc{sup −1} is an adequate fit to extraplanar gas features. The tilted ring models also present evidence for a decrease in the magnitude of the lag outside R {sub 25}, and a radial inflow of ∼10 km s{sup −1}. We extracted lagging extraplanar gas through Gaussian velocity profile fitting. From both the 3D models and extraction analyses we conclude that ∼10%–20% of the total H i mass is extraplanar. Most of the extraplanar gas is spatially coincident with regions of star formation in spiral arms, as traced by H α and GALEX FUV images, so it is likely due to star formation processes driving a galactic fountain. We also find the signature of a filament of a kinematically “forbidden” H i feature, containing ∼1.4 × 10{sup 6} M {sub ⊙} of H i, and discuss its potential relationship to a nearby H i hole. We discover a previously undetected dwarf galaxy in H i located ∼0.°4 (∼58 kpc) from the center of NGC 4559, containing ∼4 × 10{sup 5} M {sub ⊙}. This dwarf has counterpart sources in SDSS with spectra typical of H ii regions, and we conclude that it is two merging blue compact dwarf galaxies.

79 ASTRONOMY AND ASTROPHYSICS↗

Active deep kernel learning of molecular properties from structural embeddings

As vast databases of chemical identities become increasingly available, the challenge shifts to how we effectively explore and leverage these resources to study molecular properties. This paper presents an active learning approach for molecular discovery using deep kernel learning (DKL), demonstrated on the QM9 dataset. DKL links structural embeddings directly to properties, creating organized latent spaces that prioritize relevant property information. By iteratively recalculating embedding vectors in alignment with target properties, DKL uncovers concentrated maxima representing key molecular properties and reveals unexplored regions with potential for innovation. This approach underscores DKL’s potential in advancing molecular research and discovery.

Artificial neural networks↗

A survey of unsupervised learning methods for high-dimensional uncertainty quantification in black-box-type problems

Constructing surrogate models for uncertainty quantification (UQ) on complex partial differential equations (PDEs) having inherently high-dimensional O(10 n ), n ≥ 2, stochastic inputs (e.g., forcing terms, boundary conditions, initial conditions) poses tremendous challenges. The “curse of dimensionality” can be addressed with suitable unsupervised learning techniques used as a pre-processing tool to encode inputs onto lower-dimensional subspaces while retaining its structural information and meaningful properties. In this work, we review and investigate thirteen dimension reduction methods including linear and nonlinear, spectral, blind source separation, convex and non-convex methods and utilize the resulting embeddings to construct a mapping to quantities of interest via polynomial chaos expansions (PCE). Here, we refer to the general proposed approach as manifold PCE (m-PCE), where manifold corresponds to the latent space resulting from any of the studied dimension reduction methods. To investigate the capabilities and limitations of these methods we conduct numerical tests for three physics-based systems (treated as black-boxes) having high-dimensional stochastic inputs of varying complexity modeled as both Gaussian and non-Gaussian random fields to investigate the effect of the intrinsic dimensionality of input data. We demonstrate both the advantages and limitations of the unsupervised learning methods and we conclude that a suitable m-PCE model provides a cost-effective approach compared to alternative algorithms proposed in the literature, including recently proposed expensive deep neural network-based surrogates and can be readily applied for high-dimensional UQ in stochastic PDEs.

42 ENGINEERING↗

Yet Another Discriminant Analysis (YADA): A Probabilistic Model for Machine Learning Applications

This paper presents a probabilistic model for various machine learning (ML) applications. While deep learning (DL) has produced state-of-the-art results in many domains, DL models are complex and over-parameterized, which leads to high uncertainty about what the model has learned, as well as its decision process. Further, DL models are not probabilistic, making reasoning about their output challenging. In contrast, the proposed model, referred to as Yet Another Discriminate Analysis(YADA), is less complex than other methods, is based on a mathematically rigorous foundation, and can be utilized for a wide variety of ML tasks including classification, explainability, and uncertainty quantification. YADA is thus competitive in most cases with many state-of-the-art DL models. Ideally, a probabilistic model would represent the full joint probability distribution of its features, but doing so is often computationally expensive and intractable. Hence, many probabilistic models assume that the features are either normally distributed, mutually independent, or both, which can severely limit their performance. YADA is an intermediate model that (1) captures the marginal distributions of each variable and the pairwise correlations between variables and (2) explicitly maps features to the space of multivariate Gaussian variables. Numerous mathematical properties of the YADA model can be derived, thereby improving the theoretic underpinnings of ML. Validation of the model can be statistically verified on new or held-out data using native properties of YADA. However, there are some engineering and practical challenges that we enumerate to make YADA more useful.

97 MATHEMATICS AND COMPUTING↗

Performance Bounds on Two Concatenated, Interleaved Codes

A method has been developed of computing bounds on the performance of a code comprised of two linear binary codes generated by two encoders serially concatenated through an interleaver. Originally intended for use in evaluating the performances of some codes proposed for deep-space communication links, the method can also be used in evaluating the performances of short-block-length codes in other applications. The method applies, more specifically, to a communication system in which following processes take place: At the transmitter, the original binary information that one seeks to transmit is first processed by an encoder into an outer code (Co) characterized by, among other things, a pair of numbers (n,k), where n (n > k)is the total number of code bits associated with k information bits and n k bits are used for correcting or at least detecting errors. Next, the outer code is processed through either a block or a convolutional interleaver. In the block interleaver, the words of the outer code are processed in blocks of I words. In the convolutional interleaver, the interleaving operation is performed bit-wise in N rows with delays that are multiples of B bits. The output of the interleaver is processed through a second encoder to obtain an inner code (Ci) characterized by (ni,ki). The output of the inner code is transmitted over an additive-white-Gaussian- noise channel characterized by a symbol signal-to-noise ratio (SNR) Es/No and a bit SNR Eb/No. At the receiver, an inner decoder generates estimates of bits. Depending on whether a block or a convolutional interleaver is used at the transmitter, the sequence of estimated bits is processed through a block or a convolutional de-interleaver, respectively, to obtain estimates of code words. Then the estimates of the code words are processed through an outer decoder, which generates estimates of the original information along with flags indicating which estimates are presumed to be correct and which are found to be erroneous. From the perspective of the present method, the topic of major interest is the performance of the communication system as quantified in the word-error rate and the undetected-error rate as functions of the SNRs and the total latency of the interleaver and inner code. The method is embodied in equations that describe bounds on these functions. Throughout the derivation of the equations that embody the method, it is assumed that the decoder for the outer code corrects any error pattern of t or fewer errors, detects any error pattern of s or fewer errors, may detect some error patterns of more than s errors, and does not correct any patterns of more than t errors. Because a mathematically complete description of the equations that embody the method and of the derivation of the equations would greatly exceed the space available for this article, it must suffice to summarize by reporting that the derivation includes consideration of several complex issues, including relationships between latency and memory requirements for block and convolutional codes, burst error statistics, enumeration of error-event intersections, and effects of different interleaving depths. In a demonstration, the method was used to calculate bounds on the performances of several communication systems, each based on serial concatenation of a (63,56) expurgated Hamming code with a convolutional inner code through a convolutional interleaver. The bounds calculated by use of the method were compared with results of numerical simulations of performances of the systems to show the regions where the bounds are tight (see figure).

Moision, Bruce↗

Inversion of Magnetic Measurements of the Swarm A Satellite of the Bangui Magnetic Anomaly

We wanted to make a satellite altitude magnetic anomaly map of the large magnetic anomaly in the Central African Republic, the Bangui magnetic anomaly, with data from the Swarm satellites. In the first part of our study, we summarize the earlier investigations and their interpretation. In the second we discuss our data processing applied to produce a magnetic anomaly map. We used the IGRF 12th to remove the long-wavelength regional anomalies. We will use an inverse procedure, which always requires a solution of the direct problem, and a horizontal polygonal prism given in the Descartes coordinate system. For this, reason the total magnetic anomaly was transformed into the Descartes coordinate system. The magnetization and its direction were used from our previous paper. The inverse problem is solved by the Simplex procedure. Our selected polygon has 14 geometrical parameters however, the inverse problem that is the numerical determination of the minimum problem is solved in the 14 dimensions. The result of our inverse problem was the 12 horizontal coordinates and the two upper and lower data of the polygon. The origin of the Bangui anomaly has been discussed in several scientific reports, either as a deep crustal tectonic feature or the result of a large external impactor. However, according to our inversion computations we cannot make any unambiguous finding for the origin of this feature. The inaccuracy in our total anomaly map is given by the Gaussian error propagation.

KI Kis↗

Frozen Gaussian series representation of the imaginary time propagator theory and numerical tests

Thawed Gaussian wavepackets have been used in recent years to compute approximations to the thermal density matrix. From a numerical point of view, it is cheaper to employ frozen Gaussian wavepackets. In this paper, we provide the formalism for the computation of thermal densities using frozen Gaussian wavepackets. We show that the exact density may be given in terms of a series, in which the zeroth order term is the frozen Gaussian. A numerical test of the methodology is presented for deep tunneling in the quartic double well potential. In all cases, the series is observed to converge. The convergence of the diagonal density matrix element is much faster than that of the antidiagonal one, suggesting that the methodology should be especially useful for the computation of partition functions. As a by product of this study, we find that the density matrix in configuration space can have more than two saddle points at low temperatures. This has implications for the use of the quantum instanton theory.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

Intrinsic transverse momentum and parton correlations from dynamical chiral symmetry breaking

The dynamical breaking of chiral symmetry in QCD is caused by nonperturbative interactions on a distance scale rho ~ 0.3 fm, much smaller than the typical hadronic size R ~ 1 fm. These short-distance interactions influence the intrinsic transverse momentum distributions of partons and their correlations at a low normalization point. We study this phenomenon in an effective description of the low-energy dynamics in terms of chiral constituent quark degrees of freedom, which refers to the large-N_c limit of QCD. The nucleon is obtained as a system of constituent quarks and antiquarks moving in a self-consistent classical chiral field (relativistic mean-field approximation, or chiral quark-soliton model). The calculated transverse momentum distributions of constituent quarks and antiquarks are matched with QCD quarks, antiquarks and gluons at the chiral symmetry--breaking scale rho^{-2}. We find that the transverse momentum distribution of valence quarks is localized at p_T^2 ~ R^{-2} and roughly of Gaussian shape. The distribution of unpolarized sea quarks exhibits a would-be power-like tail ~1/p_T^2 extending up to the chiral symmetry-breaking scale. Similar behavior is observed in the flavor-nonsinglet polarized sea. The high-momentum tails are the result of short-range correlations between sea quarks in the nucleon&#x27;s light-cone wave function, which are analogous to short-range NN correlations in nuclei. We show that the nucleon&#x27;s light-cone wave function contains correlated pairs of transverse size rho << R with scalar-isoscalar (Sigma) and pseudoscalar-isovector (Pi) quantum numbers, whose internal wave functions have a distinctive spin structure and become identical at p_T^2 ~ rho^{-2} (restoration of chiral symmetry). These features are model-independent and represent an effect of dynamical chiral symmetry breaking on the nucleon&#x27;s partonic structure. Our results have numerous implications for the transverse momentum distributions of particles produced in hard scattering processes. Under certain conditions the nonperturbative parton correlations predicted here could be observed in particle correlations between the current and target fragmentation regions of deep-inelastic scattering.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

Unraveling the impact of initial choices and in-loop interventions on learning dynamics in autonomous scanning probe microscopy

The current focus in Autonomous Experimentation (AE) is on developing robust workflows to conduct the AE effectively. This entails the need for well-defined approaches to guide the AE process, including strategies for hyperparameter tuning and high-level human interventions within the workflow loop. This paper presents a comprehensive analysis of the influence of initial experimental conditions and in-loop interventions on the learning dynamics of Deep Kernel Learning (DKL) within the realm of AE in scanning probe microscopy. We explore the concept of the “seed effect,” where the initial experiment setup has a substantial impact on the subsequent learning trajectory. Additionally, we introduce an approach of the seed point interventions in AE allowing the operator to influence the exploration process. Using a dataset from Piezoresponse Force Microscopy on PbTiO 3 thin films, we illustrate the impact of the “seed effect” and in-loop seed interventions on the effectiveness of DKL in predicting material properties. The study highlights the importance of initial choices and adaptive interventions in optimizing learning rates and enhancing the efficiency of automated material characterization. This work offers valuable insights into designing more robust and effective AE workflows in microscopy with potential applications across various characterization techniques.

47 OTHER INSTRUMENTATION↗

How to COAAD Images. II. A Coaddition Image that is Optimal for Any Purpose in the Background-dominated Noise Limit

Image coaddition is one of the most basic operations that astronomers perform. In Paper I, we presented the optimal ways to coadd images in order to detect faint sources and to perform flux measurements under the assumption that the noise is approximately Gaussian. Here, we build on these results and derive from first principles a coaddition technique that is optimal for any hypothesis testing and measurement (e.g., source detection, flux or shape measurements, and star/galaxy separation), in the background-noise-dominated case. This method has several important properties. The pixels of the resulting coadded image are uncorrelated. This image preserves all the information (from the original individual images) on all spatial frequencies. Any hypothesis testing or measurement that can be done on all the individual images simultaneously, can be done on the coadded image without any loss of information. The PSF of this image is typically as narrow, or narrower than the PSF of the best image in the ensemble. Moreover, this image is practically indistinguishable from a regular single image, meaning that any code that measures any property on a regular astronomical image can be applied to it unchanged. In particular, the optimal source detection statistic derived in Paper I is reproduced by matched filtering this image with its own PSF. This coaddition process, which we call proper coaddition, can be understood as the maximum signal-to-noise ratio measurement of the Fourier transform of the image, weighted in such a way that the noise in the entire Fourier domain is of equal variance. This method has important implications for multi-epoch seeing-limited deep surveys, weak lensing galaxy shape measurements, and diffraction-limited imaging via speckle observations. The last topic will be covered in depth in future papers. We provide an implementation of this algorithm in MATLAB.

79 ASTRONOMY AND ASTROPHYSICS↗

Large Eddy Simulation of Cirrus Clouds

The Regional Atmospheric Modeling System (RAMS) with mesoscale interactive nested-grids and a Large-Eddy Simulation (LES) version of RAMS, coupled to two-moment microphysics and a new two-stream radiative code were used to investigate the dynamic, microphysical, and radiative aspects of the November 26, 1991 cirrus event. Wu (1998) describes the results of that research in full detail and is enclosed as Appendix 1. The mesoscale nested grid simulation successfully reproduced the large scale circulation as compared to the Mesoscale Analysis and Prediction System's (MAPS) analyses and other observations. Three cloud bands which match nicely to the three cloud lines identified in an observational study (Mace et al., 1995) are predicted on Grid #2 of the nested grids, even though the mesoscale simulation predicts a larger west-east cloud width than what was observed. Large-eddy simulations (LES) were performed to study the dynamical, microphysical, and radiative processes in the 26 November 1991 FIRE 11 cirrus event. The LES model is based on the RAMS version 3b developed at Colorado State University. It includes a new radiation scheme developed by Harrington (1997) and a new subgrid scale model developed by Kosovic (1996). The LES model simulated a single cloud layer for Case 1 and a two-layer cloud structure for Case 2. The simulations demonstrated that latent heat release can play a significant role in the formation and development of cirrus clouds. For the thin cirrus in Case 1, the latent heat release was insufficient for the cirrus clouds to become positively buoyant. However, in some special cases such as Case 2, positively buoyant cells can be embedded within the cirrus layers. These cells were so active that the rising updraft induced its own pressure perturbations that affected the cloud evolution. Vertical profiles of the total radiative and latent heating rates indicated that for well developed, deep, and active cirrus clouds, radiative cooling and latent heating could be comparable in magnitude in the cloudy layer. This implies that latent heating cannot be neglected in the construction of a cirrus cloud model. The probability density function (PDF) of w was analyzed to assist in the parameterization of cloud-scale velocities in large-scale models. For the more radiatively-driven, thin cirrus case, the PDFs are approximately Gaussian. However, in the interior of the deep, convectively unstable case, the PDFs of w are multi-modal and very broad, indicating that parameterizing cloud-scale motions for such clouds can be very challenging. The results of this research are described in detail in a paper submitted to the Journal of Atmospheric Science (Wu and Cotton, 1999), which is enclosed as Appendix 2. Using soundings extracted from a mesoscale simulation of the November 26, 1991 cirrus event, the radiative effects on vapor deposition/sublimation of ice crystals was studied using a two-dimensional cloud-resolving model (CRM) version of RAMS, coupled to an explicit bin-resolving microphysics. The CRM simulations of the November 26, 1991 cirrus event demonstrate that the radiative impact on the diffusional growth (or sublimation) of ice crystals is significant. In this case, the ice particles experienced radiative warming. Model results show that radiative feedbacks in the diffusional growth of ice particles can be very complex. Radiative warming of an ice particle will restrict the particle's diffusional growth. In the case of radiative warming, ice particles larger than a certain size will experience so much radiative warming that surface ice saturation vapor pressures become large enough to cause sublimation of the larger crystals, while smaller crystals are growing by vapor deposition. However, ice mass production can be enhanced in the case of radiative cooling of an ice particle. For the November 26, 1991 cirrus event, radiative feedback results in significant reduction in the total ice mass, especially in the production of large ice crystals, and consequently, both radiative and dynamic properties of the cirrus cloud are significantly affected. A complete description of this research has been submitted as a paper to the Journal of Atmospheric Science (Wu et al., 1999), and included as Appendix 3.

Wu, Ting↗

Less is More: Bigger Data from Compressive Measurements

Compressive sensing approaches are beginning to take hold in (scanning) transmission electron microscopy (S/TEM) [1,2,3]. Compressive sensing is a mathematical theory about acquiring signals in a compressed form (measurements) and the probability of recovering the original signal by solving an inverse problem [4]. The inverse problem is underdetermined (more unknowns than measurements), so it is not obvious that recovery is possible. Compression is achieved by taking inner products of the signal with measurement weight vectors. Both Gaussian random weights and Bernoulli (0,1) random weights form a large class of measurement vectors for which recovery is possible. The measurements can also be designed through an optimization process. The key insight for electron microscopists is that compressive sensing can be used to increase acquisition speed and reduce dose. Building on work initially developed for optical cameras, this new paradigm will allow electron microscopists to solve more problems in the engineering and life sciences. We will be collecting orders of magnitude more data than previously possible. The reason that we will have more data is because we will have increased temporal/spatial/spectral sampling rates, and we will be able ability to interrogate larger classes of samples that were previously too beam sensitive to survive the experiment. For example consider an in-situ experiment that takes 1 minute. With traditional sensing, we might collect 5 images per second for a total of 300 images. With compressive sensing, each of those 300 images can be expanded into 10 more images, making the collection rate 50 images per second, and the decompressed data a total of 3000 images [3]. But, what are the implications, in terms of data, for this new methodology? Acquisition of compressed data will require downstream reconstruction to be useful. The reconstructed data will be much larger than traditional data, we will need space to store the reconstructions during analysis, and the computational demands for analysis will be higher. Moreover, there will be time costs associated with reconstruction. Deep learning [5] is an approach to address these problems. Deep learning is a hierarchical approach to find useful (for a particular task) representations of data. Each layer of the hierarchy is intended to represent higher levels of abstraction. For example, a deep model of faces might have sinusoids, edges and gradients in the first layer; eyes, noses, and mouths in the second layer, and faces in the third layer. There has been significant effort recently in deep learning algorithms for tasks beyond image classification such as compressive reconstruction [6] and image segmentation [7]. A drawback of deep learning, however, is that training the model requires large datasets and dedicated computational resources (to reduce training time to a few days). A second issue is that deep learning is not user-friendly and the meaning behind the results is usually not interpretable. We have shown it is possible to reduce the data set size while maintaining model quality [8] and developed interpretable models for image classification [9], but the demands are still significant. The key to addressing these problems is to NOT reconstruct the data. Instead, we should design computational sensors that give answers to specific problems. A simple version of this idea is compressive classification [10], where the goal is to classify signal type from a small number of compressed measurements. Classification is a much simpler problem than reconstruction, so 1) much fewer measurements will be necessary, and 2) these measurements will probably not be useful for reconstruction. Other simple examples of computational sensing include determining object volume or the number of objects present in the field of view [11].

Stevens, Andrew↗

Triton: Igiugig Fish Video Analysis (Project Report)

This document summarizes the analysis of video data collected around the Ocean Renewable Power Company’s RivGen® device in the Kvichak River during July and August 2015. Human (manual) analysis of the data was undertaken and used to develop automated algorithms to detect fish in the frames and describe interaction behavior. In addition, a web application, EyeSea, was developed to combine manual and automated processing, so that ultimately the automated algorithms could be used to identify where human analysis was needed (i.e., when fish are present in video frames). The aim of the project was to develop algorithms that could identify video frames with fish present that could then be used to enable faster manual analysis. However, the project has also identified recommendations for future data collection, and research that could help analysis and interpretation. The manual analysis started to look at all data from the start of the deployment; this was to ensure rare events were seen, and initially focused on night-time data when more fish were present. This process highlighted the amount of time taken to identify fish, and ultimately only 42.33 hours of video were reviewed due to the time-consuming analysis. The data were classified as ‘Fish’, when the reviewer was confident it was a fish, and ‘Maybe’ fish, when it was difficult to distinguish. This was based on the movement, shape and color characteristics. ‘Fish’ were further classified as ‘adult’, ‘juvenile’, or ‘unidentifiable’. Behavioral attributes were noted, which were broadly divided into passive and avoidance activities. In over 42 hours data, there were only 20 potential contact interactions, of which 3 were ‘Maybe’ classifications, 12 were juveniles and 5 were adults; on only one occasion was an actual contact confirmed, and this was an adult fish with the camera, not the turbine itself. This also highlights the difficulties associated with confirming a strike or collision event as occurring or being a near-miss. More interactions were detected at night; this was probably biased by nighttime artificial light use, which may have attracted fish, but also could have increased detection probability as the light is reflected from the fish itself. For the algorithm development, background subtraction, optical flow, and deep learning techniques were considered. The deep learning approach was determined as needing too much training data for this application, and therefore was not continued. The optical flow analysis was considered promising, but did not give immediate results, so would need further investigation. Therefore, background subtraction was the main focus. Three methods of background subtraction were tried: Robust Principal Components Analysis (RPCA), Gaussian Mixture Model (GMM) and Video Background Extraction(ViBE). A classification technique was then applied to the foreground images to determine fish presence. Using this combination, it was found that fish could be accurately detected when occupying more pixels (>200 pixels, 98.2% correct; 100-200 pixels, 99.6% correct; 5-100 pixels, 85.4% correct; 2-5 pixels, 66.3% correct). In parallel, EyeSea was developed to convert the video data to a usable form and to enable manual and automated analysis of the data that would have a standardized output. Recommendations for further research, and standardized methods for data collection are given.

13 HYDRO ENERGY↗