Ninan Sajeeth Philip

Results from the Supernova Photometric Classification Challenge

We report results from the Supernova Photometric Classification Challenge (SNPhotCC), a publicly released mix of simulated supernovae (SNe), with types (Ia, Ibc, and II) selected in proportion to their expected rates. The simulation was realized in the griz filters of the Dark Energy Survey (DES) with realistic observing conditions (sky noise, point-spread function, and atmospheric transparency) based on years of recorded conditions at the DES site. Simulations of non-Ia-type SNe are based on spectroscopically confirmed light curves that include unpublished non-Ia samples donated from the Carnegie Supernova Project (CSP), the Supernova Legacy Survey (SNLS), and the Sloan Digital Sky Survey-II (SDSS-II). A spectroscopically confirmed subset was provided for training. We challenged scientists to run their classification algorithms and report a type and photo-z for each SN. Participants from 10 groups contributed 13 entries for the sample that included a host-galaxy photo-z for each SN and nine entries for the sample that had no redshift information. Several different classification strategies resulted in similar performance, and for all entries the performance was significantly better for the training subset than for the unconfirmed sample. For the spectroscopically unconfirmed subset, the entry with the highest average figure of merit for classifying SNe Ia has an efficiency of 0.96 and an SN Ia purity of 0.79. As a public resource for the future development of photometric SN classification and photo-z estimators, we have released updated simulations with improvements based on our experience from the SNPhotCC, added samples corresponding to the Large Synoptic Survey Telescope (LSST) and the SDSS-II, and provided the answer keys so that developers can evaluate their own analysis.

Automated Galaxy Morphology: A Fourier Approach

We use automated surface photometry and pattern classification techniques to morphologically classify galaxies. The two-dimensional light distribution of a galaxy is reconstructed using Fourier series fits to azimuthal profiles computed in concentric elliptical annuli centered on the galaxy. Both the phase and amplitude of each Fourier component have been studied as a function of radial bin number for a large collection of galaxy images using principal-component analysis. We find that up to 90% of the variance in many of these Fourier profiles may be characterized in as few as three principal components and that their use substantially reduces the dimensionality of the classification problem. We use supervised learning methods in the form of artificial neural networks to train galaxy classifiers that detect morphological bars at the 85%-90% confidence level and can identify the Hubble type with a 1 σ scatter of 1.5 steps on the 16 step stage axis of the revised Hubble system. Finally, we systematically characterize the adverse effects of decreasing resolution and signal-to-noise ratio on the quality of morphological information predicted by these classifiers.

A neural network tool for analyzing trends in rainfall

Rainfall, like all other natural phenomena is highly unpredictable. Traditionally, principal component analysis and spectral analysis are used to understand trends in rainfall over long periods. In this paper, we present a new method that appears to be better than existing methods to understand the long-term behavior of rainfall phenomena. Using a case study on the rainfall in Kerala State, the southern part of Indian Peninsula, we show that a new kind of neural network known as the adaptive basis function network is a promising tool for climatic studies, especially rainfall analysis. The paper also reveals that in spite of the fluctuations resulting from the nonlinearity in the system, the trends in the rainfall pattern in Kerala state have remained unaffected over the past 87 years from 1893 to 1980. We also successfully filter out the chaotic part of the system and illustrate that its effects are marginal over long-term predictions.

A Difference boosting neural network for automated star-galaxy classification

In this paper we describe the use of a new articial neural network, called the dierence boosting neural network (DBNN), for automated classication problems in astronomical data analysis. We illustrate the capabilities of the network by applying it to star galaxy classication using recently released, deep imaging data. We have compared our results with classication made by the widely used Source Extractor (SExtractor) package. We show that while the performance of the DBNN in star-galaxy classication is comparable to that of SExtractor, it has the advantage of signicantly higher speed and flexibility during training as well as classication.

Transient Classification in LIGO data using Difference Boosting Neural Network

Detection and classification of transients in data from gravitational wave detectors are crucial for efficient searches for true astrophysical events and identification of noise sources. We present a hybrid method for classification of short duration transients seen in gravitational wave data using both supervised and unsupervised machine learning techniques. To train the classifiers we use the relative wavelet energy and the corresponding entropy obtained by applying one-dimensional wavelet decomposition on the data. The prediction accuracy of the trained classifier on 9 simulated classes of gravitational wave transients and also LIGOs sixth science run hardware injections are reported. Targeted searches for a couple of known classes of non-astrophysical signals in the first observational run of Advanced LIGO data are also presented. The ability to accurately identify transient classes using minimal training samples makes the proposed method a useful tool for LIGO detector characterization as well as searches for short duration gravitational wave signals.

A wavelet based algorithm for the identification of oscillatory event-related potential components.

Event related potentials (ERPs) are very feeble alterations in the ongoing electroencephalogram (EEG) and their detection is a challenging problem. Based on the unique time-based parameters derived from wavelet coefficients and the asymmetry property of wavelets a novel algorithm to separate ERP components in single-trial EEG data is described. Though illustrated as a specific application to N170 ERP detection, the algorithm is a generalized approach that can be easily adapted to isolate different kinds of ERP components. The algorithm detected the N170 ERP component with a high level of accuracy. We demonstrate that the asymmetry method is more accurate than the matching wavelet algorithm and t-CWT method by 48.67 and 8.03 percent, respectively. This paper provides an off-line demonstration of the algorithm and considers issues related to the extension of the algorithm to real-time applications.

Adaptive basis function for artificial neural networks

Abstract It is shown that modifying the sigmoidal basis function of a multi-layer feedforward artificial neural network using a control parameter improves the networks ability to learn. The modification is rendered by a gradient descent algorithm similar to the back-propagation. In doing so, the method retains all the goodies of the sigmoidal function and causes the ANN to approximate the decision function faster and also with better accuracy.

Determining the torus covering factors for a sample of type 1 AGN in the local Universe.

In the unified scheme of active galactic nuclei, a dusty torus absorbs and then reprocesses a fraction of the intrinsic luminosity which is emitted at longer wavelengths. Thus, subject to radiative transfer corrections, the fraction of the sky covered by the torus as seen from the central source (known as the covering factor fc) can be estimated from the ratio of the infrared to the bolometric luminosities of the source as fc = Ltorus/LBol. However, the uncertainty in determining LBol has made the estimation of covering factors by this technique difficult, especially for AGN in the local Universe where the peak of the observed spectral energy distributions lies in the UV (ultraviolet). Here, we determine the covering factors of an X-ray/optically selected sample of 51 type 1 AGN. The bolometric luminosities of these sources are derived using a self-consistent, energy-conserving model that estimates the contribution in the unobservable far-UV region, using multifrequency data obtained from SDSS, XMM–Newton, WISE, 2MASS and UKIDSS. We derive a mean value of fc ∼ 0.30 with a dispersion of 0.17. Sample correlations, combined with simulations, show that fc is more strongly anticorrelated with λEdd than with LBol. This points to large-scale torus geometry changes associated with the Eddington-dependent accretion flow, rather than a receding torus, with its inner sublimation radius determined solely by heating from the central source. Furthermore, we do not see any significant change in the distribution of fc for sub-samples of radio-loud sources or Narrow Line Seyfert 1 galaxies (NLS1s), though these sub-samples are small.

What is there in a training sample

Two factors that are known to have direct influence on the classification accuracy of any neural network are (1) the network complexity and (2) the representational accuracy of the training data. While pruning algorithms are used to tackle the complexity problem, no direct solutions are known for the second. Selecting training data at random from the sample space is the most popular method followed. Despite its simplicity, this method does not ensure nor guarantee that the training would be optimal. In this brief paper, we present a new method that is specific to a difference boosting neural network (DBNN) but could probably be extended to other networks as well. The method is iterative and fast, ensuring optimal selection of the minimum training data from a larger set in an automated manner. We test the performance of the new method on the some of the well known datasets from the UCI repository for benchmarking machine learning tools and show that the performance of the new method in almost all cases is better than that in any published method of comparable network complexity and that it requires only a fraction of the usual training data, thereby, making learning faster and more generic.

Detection of bars in galaxies using a deep convolutional neural network

We present an efficient and automated method for the detection of bar structure in optical images of galaxies using a deep convolutional neural network. In our study we use a sample of 9346 galaxies in the redshift range 0.009-0.2 from the Sloan Digital Sky Survey, which has 3864 barred galaxies, the rest being unbarred. We reach a top precision of ~94 per cent in identifying bars in galaxies using the trained network. This accuracy matches the accuracy reached by human experts on the same data without additional information about the images. The method can be easily scaled to the much larger samples anticipated from various surveys.