Neda Rohani

Gravity Spy: Integrating Advanced LIGO Detector Characterization, Machine Learning, and Citizen Science

With the first direct detection of gravitational waves, the advanced laser interferometer gravitational-wave observatory (LIGO) has initiated a new field of astronomy by providing an alternative means of sensing the universe. The extreme sensitivity required to make such detections is achieved through exquisite isolation of all sensitive components of LIGO from non-gravitational-wave disturbances. Nonetheless, LIGO is still susceptible to a variety of instrumental and environmental sources of noise that contaminate the data. Of particular concern are noise features known as glitches, which are transient and non-Gaussian in their nature, and occur at a high enough rate so that accidental coincidence between the two LIGO detectors is non-negligible. Glitches come in a wide range of time-frequency-amplitude morphologies, with new morphologies appearing as the detector evolves. Since they can obscure or mimic true gravitational-wave signals, a robust characterization of glitches is paramount in the effort to achieve the gravitational-wave detection rates that are predicted by the design sensitivity of LIGO. This proves a daunting task for members of the LIGO Scientific Collaboration alone due to the sheer amount of data. In this paper we describe an innovative project that combines crowdsourcing with machine learning to aid in the challenging task of categorizing all of the glitches recorded by the LIGO detectors. Through the Zooniverse platform, we engage and recruit volunteers from the public to categorize images of time-frequency representations of glitches into pre-identified morphological classes and to discover new classes that appear as the detectors evolve. In addition, machine learning algorithms are used to categorize images after being trained on human-classified examples of the morphological classes. Leveraging the strengths of both classification methods, we create a combined method with the aim of improving the efficiency and accuracy of each individual classifier. The resulting classification and characterization should help LIGO scientists to identify causes of glitches and subsequently eliminate them from the data or the detector entirely, thereby improving the rate and accuracy of gravitational-wave observations. We demonstrate these methods using a small subset of data from LIGOs first observing run.

Variational Gaussian process for sensor fusion

In this paper, we introduce a new Gaussian Process (GP) classification method for multisensory data. The proposed approach can deal with noisy and missing data. It is also capable of estimating the contribution of each sensor towards the classification task. We use Bayesian modeling to build a GP-based classifier which combines the information provided by all sensors and approximates the posterior distribution of the GP using variational Bayesian inference. During its training phase, the algorithm estimates each sensors weight and then uses this information to assign a label to each new sample. In the experimental section, we evaluate the classiication performance of the proposed method on both synthetic and real data and show its applicability to different scenarios.

Matrix sparsification and non-negative factorization for task partitioning in computational sensing and imaging

We address the mathematical foundations of a special case of the general problem of partitioning an end-to-end sensing algorithm for implementation by optics and by a digital processor for minimal electrical power dissipation. Specifically, we present a non-iterative algorithm for factoring a general k × k real matrix A (describing the end-to-end linear pre-processing) into the product BC, where C has no negative entries (for implementation in linear optics) and B is maximally sparse, i.e., has the fewest possible non-zero entries (for minimal dissipation of electrical power). Our algorithm achieves a sparsification of B: i.e., the number s of non-zero entries in B: of s ≤ 2k, which we prove is optimal for our class of problems.

Innovative data reduction and visualization strategy for hyperspectral imaging datasets using t-SNE approach

Abstract Visible hyperspectral imaging (HSI) is a fast and non-invasive imaging method that has been adapted by the field of conservation science to study painted surfaces. By collecting reflectance spectra from a 2D surface, the resulting 3D hyperspectral data cube contains millions of recorded spectra. While processing such large amounts of spectra poses an analytical and computational challenge, it also opens new opportunities to apply powerful methods of multivariate analysis for data evaluation. With the intent of expanding current data treatment of hyperspectral datasets, an innovative approach for data reduction and visualization is presented in this article. It uses a statistical embedding method known as t-distributed stochastic neighbor embedding (t-SNE) to provide a non-linear representation of spectral features in a lower 2D space. The efficiency of the proposed method for painted surfaces from cultural heritage is established through the study of laboratory prepared paint mock-ups, and medieval French illuminated manuscript.

Variational Gaussian process for multisensor classification problems

Abstract This paper proposes a new model for multi-sensory data classification. To tackle this problem, probabilistic modeling and variational Bayesian inference are used. A Gaussian Process (GP) classifier is built upon the introduced modeling. Its posterior distribution is approximated using variational Bayesian inference. Finally, labels of test samples are predicted employing this classifier. Very importantly, and in contrast to alternative approaches, the proposed method does not discard samples with missing features and utilizes all available information for training. Furthermore, to take into account that the quality of the information provided by each sensor may differ (some modalities/sensors may provide more reliable/distinctive information than others), we introduce two versions of the algorithm. In the first one, the parameters modeling each sensor performance are shared while in the second one, each sensor parameters are estimated independently. Synthetic and real datasets are utilized to examine the validity of the proposed models. The results obtained for binary classification problems justify their use and confirm their superiority over existing fusion architectures.

Nonlinear Unmixing of Hyperspectral Datasets for the Study of Painted Works of Art

Nonlinear unmixing of hyperspectral reflectance data is one of the key problems in quantitative imaging of painted works of art. The approach presented is to interrogate a hyperspectral image cube by first decomposing it into a set of reflectance curves representing pure basis pigments and second to estimate the scattering and absorption coefficients of each pigment in a given pixel to produce estimates of the component fractions. This two-step algorithm uses a deep neural network to qualitatively identify the constituent pigments in any unknown spectrum and, based on the pigment(s) present and Kubelka-Munk theory to estimate the pigment concentration on a per-pixel basis. Using hyperspectral data acquired on a set of mock-up paintings and a well-characterized illuminated folio from the 15th century, the performance of the proposed algorithm is demonstrated for pigment recognition and quantitative estimation of concentration.

Distinguishing Nigerian Food Items and Calorie Content with Hyperspectral Imaging

Identifying food types consumed and their calorie composition is one of the central tasks of dietary assessment. Traditional automated image processing methods learn to map images to an existing food database with known caloric composition. However, even when the correct food type is identified, caloric makeup can vary depending on its ingredients, and using true-color images proves insufficient to distinguish within food type variability. In this paper, we show that hyperspectral imaging provides useful information and promise in distinguishing caloric composition within the same food type. We collect data using a hyperspectral camera from Nigerian foods cooked with varying degrees of fat content, and capture images under different intensities of light. We apply Principle Component Analysis (PCA) to reduce the dimensionality, and train a Support Vector Machine (SVM) classifier using a Radial Basis Function kernel and show that applying this technique on hyperspectral images can more readily distinguish calorie composition. Furthermore, compared with methods that only use true-color based features, our method shows that a classifier trained using features from hyperspectral images is significantly more predictive of within-food caloric content, and by fusing results from two classifiers trained separately using hyperspectral and RGB imagery we obtain the greatest predictive power.

Deep multi-view models for glitch classification

Non-cosmic, non-Gaussian disturbances known as “glitches”, show up in gravitational-wave data of the Advanced Laser Interferometer Gravitational-wave Observatory, or aLIGO. In this paper, we propose a deep multi-view convolutional neural network to classify glitches automatically. The primary purpose of classifying glitches is to understand their characteristics and origin, which facilitates their removal from the data or from the detector entirely. We visualize glitches as spectrograms and leverage the state-of-the-art image classification techniques in our model. The suggested classifier is a multi-view deep neural network that exploits four different views for classification. The experimental results demonstrate that the proposed model improves the overall accuracy of the classification compared to traditional single view algorithms.

Automatic pigment identification on roman Egyptian paintings by using sparse modeling of hyperspectral images

In this paper, we study the problem of automatic identification of pigments applied to paintings using hyperspectral reflectance data. Here, we cast the problem of pigment identification in a novel way by decomposing the spectrum into pure pigments. The pure pigment exemplars, chosen and prepared in our laboratory based on historic sources and archaeological examples, closely resemble the materials used to make ancient paintings. To validate our algorithm, we created a set of mock-up paintings in our laboratory consisting of a broad palette of mixtures of pure pigments. Our results clearly demonstrate more accurate estimation of pigment composition than purely distance-based methods such as spectral angle mapping (SAM) and spectral correlation mapping (SCM). In addition, we studied hyperspectral imagery acquired of a Roman-Egyptian portrait, excavated from the site of Tebtunis in the Fayum region of Egypt, and dated to about the 2nd century CE. Using ground truth information obtained using Raman spectroscopy, we show qualitatively that our method accurately detects pigment composition for the specific pigments hematite and indigo.