arXiv subjects

Masayuki Tanaka

Publications and source records attributed to Masayuki Tanaka.

At least 109 records · Page 6Linked to original sources

arXiv2019

New Perspective of Interpretability of Deep Neural Networks

Deep neural networks (DNNs) are known as black-box models. In other words, it is difficult to interpret the internal state of the model. Improving the interpretability of DNNs is one of the hot research topics. However, at present, the definition of interpretability for DNNs is vague, and the question of what is a highly explanatory model is still controversial. To address this issue, we provide the definition of the human predictability of the model, as a part of the interpretability of the DNNs. The human predictability proposed in this paper is defined by easiness to predict the change of the inference when perturbating the model of the DNNs. In addition, we introduce one example of high human-predictable DNNs. We discuss that our definition will help to the research of the interpretability of the DNNs considering various types of applications.

cs.LG↗

arXiv2019

The CFHT Large Area U-band Deep Survey (CLAUDS)

The CFHT Large Area U-band Deep Survey (CLAUDS) uses data taken with the MegaCam mosaic imager on CFHT to produce images of 18.60 deg2 with median seeing of FWHM=0.92 arcsec and to a median depth of U = 27.1 AB (5 sigma in 2 arcsec apertures), with selected areas that total 1.36 deg2 reaching a median depth of U=27.7 AB. These are the deepest U-band images assembled to date over this large an area. These data are located in four fields also imaged to comparably faint levels in grizy and several narrowband filters as part of the Hyper Suprime-Cam (HSC) Subaru Strategic Program (HSC-SSP). These CFHT and Subaru datasets will remain unmatched in their combination of area and depth until the advent of the Large Synoptic Survey Telescope (LSST). This paper provides an overview of the scientific motivation for CLAUDS and gives details of the observing strategy, observations, data reduction, and data merging with the HSC-SSP. Three early applications of these deep data are used to illustrate the potential of the dataset: deep U-band galaxy number counts, z~3 Lyman break galaxy (LBG) selection, and photometric redshifts improved by adding CLAUDS U to the Subaru HSC grizy photometry.

astro-ph.GA↗

arXiv2019

Second Data Release of the Hyper Suprime-Cam Subaru Strategic Program

This paper presents the second data release of the Hyper Suprime-Cam Subaru Strategic Program, a wide-field optical imaging survey on the 8.2 meter Subaru Telescope. The release includes data from 174 nights of observation through January 2018. The Wide layer data cover about 300 deg^2 in all five broadband filters (grizy) to the nominal survey exposure (10min in gr and 20min in izy). Partially observed areas are also included in the release; about 1100 deg^2 is observed in at least one filter and one exposure. The median seeing in the i-band is 0.6 arcsec, demonstrating the superb image quality of the survey. The Deep (26 deg^2) and UltraDeep (4 deg^2) data are jointly processed and the UltraDeep-COSMOS field reaches an unprecedented depth of i~28 at 5 sigma for point sources. In addition to the broad-bands, narrow-band data are also available in the Deep and UltraDeep fields. This release includes a major update to the processing pipeline, including improved sky subtraction, PSF modeling, object detection, and artifact rejection. The overall data quality has been improved, but this release is not without problems; there is a persistent deblender problem as well as new issues with masks around bright stars. The user is encouraged to review the issue list before utilizing the data for scientific explorations. All the image products as well as catalog products are available for download. The catalogs are also loaded to a database, which provides an easy interface for users to retrieve data for objects of interest. In addition to these main data products, detailed galaxy shape measurements withheld from the Public Data Release 1 (PDR1) are now available to the community. The shape catalog is drawn from the S16A internal release, which has a larger area than PDR1 (160 deg^2). All products are available at the data release site, https://hsc-release.mtk.nao.ac.jp/.

astro-ph.IM↗

arXiv2019

Subaru High-z Exploration of Low-Luminosity Quasars (SHELLQs). X. Discovery of 35 Quasars and Luminous Galaxies at 5.7 $\le$ z $\le$ 7.0

We report the discovery of 28 quasars and 7 luminous galaxies at 5.7 $\le$ z $\le$ 7.0. This is the tenth in a series of papers from the Subaru High-z Exploration of Low-Luminosity Quasars (SHELLQs) project, which exploits the deep multi-band imaging data produced by the Hyper Suprime-Cam (HSC) Subaru Strategic Program survey. The total number of spectroscopically identified objects in SHELLQs has now grown to 93 high-z quasars, 31 high-z luminous galaxies, 16 [O III] emitters at z ~ 0.8, and 65 Galactic cool dwarfs (low-mass stars and brown dwarfs). These objects were found over 900 deg2, surveyed by HSC between 2014 March and 2018 January. The full quasar sample includes 18 objects with very strong and narrow Ly alpha emission, whose stacked spectrum is clearly different from that of other quasars or galaxies. While the stacked spectrum shows N V 1240 emission and resembles that of lower-z narrow-line quasars, the small Ly alpha width may suggest a significant contribution from the host galaxies. Thus these objects may be composites of quasars and star-forming galaxies.

astro-ph.GA↗

arXiv2019

Subaru High-z Exploration of Low-Luminosity Quasars (SHELLQs). VI. Black Hole Mass Measurements of Six Quasars at 6.1<z<6.7

We present deep near-infrared spectroscopy of six quasars at 6.1 5.8 quasars with measured black hole masses. From single-epoch mass measurements based on MgII2798, we find a wide range in black hole masses, from M_BH=10^7.6 to 10^9.3 Msun. The Eddington ratios L_bol/L_Edd range from 0.16 to 1.1, but the majority of the HSC quasars are powered by M_BH=10^9 Msun supermassive black holes (SMBHs) accreting at sub-Eddington rates. The Eddington ratio distribution of the HSC quasars is inclined to lower accretion rates than those of Willott et al. (2010a), who measured the black hole masses for similarly faint z=6 quasars. This suggests that the global Eddington ratio distribution is wider than has previously been thought. The presence of M_BH=10^9 Msun SMBHs at z=6 cannot be explained with constant sub-Eddington accretion from stellar remnant seed black holes. Therefore, we may be witnessing the first buildup of the most massive black holes in the first billion years of the universe, the accretion activity of which is transforming from active growth to a quiescent phase. Measurements of a larger complete sample of z>6 low-luminosity quasars, as well as deeper observations with future facilities will enable us to better understand the early SMBH growth in the reionization epoch.

astro-ph.GA↗

arXiv2019

HSC16aayt: Slowly evolving interacting transient rising for more than 100 days

We report our observations of HSC16aayt (SN 2016jiu), which was discovered by the Subaru/Hyper Suprime-Cam (HSC) transient survey conducted as part of Subaru Strategic Program (SSP). It shows very slow photometric evolution and its rise time is more than 100 days. The optical magnitude change in 400 days remains within 0.6 mag. Spectra of HSC16aayt show a strong narrow emission line and we classify it as a Type IIn supernova. The redshift of HSC16aayt is 0.6814 +/- 0.0002 from the spectra. Its host galaxy center is at 5 kpc from the supernova location and HSC16aayt might be another example of isolated Type IIn supernovae, although the possible existence of underlying star forming activity of the host galaxy at the supernova location is not excluded.

astro-ph.HE↗

arXiv2019

Boötes IV: A New Milky Way Satellite Discovered in the Subaru Hyper Suprime-Cam Survey and Implications for the Missing Satellite Problem

We report on the discovery of a new Milky Way (MW) satellite in Boötes based on data from the on-going Hyper Suprime-Cam (HSC) Subaru Strategic Program (SSP). This satellite, named Boötes IV, is the third ultra-faint dwarf that we have discovered in the HSC-SSP. We have identified a statistically significant (32.3$σ$) overdensity of stars having characteristics of a metal-poor, old stellar population. The distance to this stellar system is $D_{\odot}=209^{+20}_{-18}$ kpc with a $V$-band absolute magnitude of $M_V=-4.53^{+0.23}_{-0.21}$ mag. Boötes IV has a half-light radius of $r_h=462^{+98}_{-84}$ pc and an ellipticity of $0.64^{+0.05}_{-0.05}$, which clearly suggests that this is a dwarf satellite galaxy. We also found another overdensity that appears to be a faint globular cluster with $M_V=-0.20^{+0.59}_{-0.83}$ mag and $r_h=5.9^{+1.5}_{-1.3}$ pc located at $D_{\odot}=46^{+4}_{-4}$ kpc. Adopting the recent prediction for the total population of satellites in a MW-sized halo by Newton et al. (2018), which combined the characteristics of the observed satellites by SDSS and DES with the subhalos obtained in $Λ$CDM models, we estimate that there should be about two MW satellites at $M_V\le0$ in the $\sim676$ deg$^2$ covered by HSC-SSP, whereas that area includes six satellites. Thus, the observed number of satellites is larger than the theoretical prediction. On the face of it, we have a problem of too many satellites, instead of the well-known missing satellites problem whereby the $Λ$CDM theory overpredicts the number of satellites in a MW-sized halo. This may imply that the models need more refinements for the assignment of subhalos to satellites such as considering those found by the current deeper survey. [abridged]

astro-ph.GA↗

arXiv2019

The stellar halo of the Milky Way traced by blue horizontal-branch stars in the Subaru Hyper Suprime-Cam Survey

We report on the global structure of the Milky Way (MW) stellar halo up to its outer boundary based on the analysis of blue-horizontal branch stars (BHBs). These halo tracers are extracted from the $(g,r,i,z)$ band multi-photometry in the internal data release of the on-going Hyper Suprime-Cam Subaru Strategic Program (HSC-SSP) surveyed over $\sim550$~deg$^2$ area. In order to select most likely BHBs by removing blue straggler stars (BSs) and other contamination in a statistically significant manner, we have developed and applied an extensive Bayesian method, instead of the simple color cuts adopted in our previous work, where each of the template BHBs and non-BHBs obtained from the available catalogs is represented as a mixture of multiple Gaussian distributions in the color-color diagrams. We found from the candidate BHBs in the range of 18.5 r_{\rm b}$. The latter profile with a prolate shape having an axial ratio of $q=1.72^{+0.44}_{-0.28}$ is most likely and this halo may hold a rather sharp boundary at r=160kpc. The slopes of the halo density profiles are compared with those from the suite of hydrodynamical simulations for the formation of stellar halos. This comparison suggests that the MW stellar halo may consist of the two overlapping components: the in situ. inner halo as probed by RR Lyrae stars showing a relatively steep radial density profile and the ex situ. outer halo with a shallow profile probed by BHBs here, which is made by accretion of small stellar systems.

astro-ph.GA↗

arXiv2019

Breaking Inter-Layer Co-Adaptation by Classifier Anonymization

This study addresses an issue of co-adaptation between a feature extractor and a classifier in a neural network. A naive joint optimization of a feature extractor and a classifier often brings situations in which an excessively complex feature distribution adapted to a very specific classifier degrades the test performance. We introduce a method called Feature-extractor Optimization through Classifier Anonymization (FOCA), which is designed to avoid an explicit co-adaptation between a feature extractor and a particular classifier by using many randomly-generated, weak classifiers during optimization. We put forth a mathematical proposition that states the FOCA features form a point-like distribution within the same class in a class-separable fashion under special conditions. Real-data experiments under more general conditions provide supportive evidences.

cs.LG↗

arXiv2019

Intentional Attention Mask Transformation for Robust CNN Classification

Convolutional Neural Networks have achieved impressive results in various tasks, but interpreting the internal mechanism is a challenging problem. To tackle this problem, we exploit a multi-channel attention mechanism in feature space. Our network architecture allows us to obtain an attention mask for each feature while existing CNN visualization methods provide only a common attention mask for all features. We apply the proposed multi-channel attention mechanism to multi-attribute recognition task. We can obtain different attention mask for each feature and for each attribute. Those analyses give us deeper insight into the feature space of CNNs. Furthermore, our proposed attention mechanism naturally derives a method for improving the robustness of CNNs. From the observation of feature space based on the proposed attention mask, we demonstrate that we can obtain robust CNNs by intentionally emphasizing features that are important for attributes. The experimental results for the benchmark dataset show that the proposed method gives high human interpretability while accurately grasping the attributes of the data, and improves network robustness.

cs.CV↗

arXiv2019

A Wide and Deep Exploration of Radio Galaxies with Subaru HSC (WERGS). II. Physical Properties derived from the SED Fitting with Optical, Infrared, and Radio Data

We present physical properties of radio galaxies (RGs) with $f_{\rm 1.4 GHz} >$ 1 mJy discovered by Subaru Hyper Supreme-Cam (HSC) and VLA Faint Images of the Radio Sky at Twenty-Centimeters (FIRST) survey. For 1056 FIRST RGs at $0 < z \leq 1.7$ with HSC counterparts in about 100 deg$^2$, we compiled multi-wavelength data of optical, near-infrared (IR), mid-IR, far-IR, and radio (150 MHz). We derived their color excess ($E (B-V)_{*}$), stellar mass, star formation rate (SFR), IR luminosity, the ratio of IR and radio luminosity ($q_{\rm IR}$), and radio spectral index ($α_{\rm radio}$) that are derived from the SED fitting with CIGALE. We also estimated Eddington ratio based on stellar mass and integration of the best-fit SEDs of AGN component. We found that $E (B-V)_{*}$, SFR, and IR luminosity clearly depend on redshift while stellar mass, $q_{\rm IR}$, and $α_{\rm radio}$ do not significantly depend on redshift. Since optically-faint ($i_{\rm AB} \geq 21.3$) RGs that are newly discovered by our RG survey tend to be high redshift, they tend to not only have a large dust extinction and low stellar mass but also have high SFR and AGN luminosity, high IR luminosity, and high Eddington ratio compared to optically-bright ones. The physical properties of a fraction of RGs in our sample seem to differ from a classical view of RGs with massive stellar mass, low SFR, and low Eddington ratio, demonstrating that our RG survey with HSC and FIRST provides us curious RGs among entire RG population.

astro-ph.GA↗

arXiv2019

Interpretation of Feature Space using Multi-Channel Attentional Sub-Networks

Convolutional Neural Networks have achieved impressive results in various tasks, but interpreting the internal mechanism is a challenging problem. To tackle this problem, we exploit a multi-channel attention mechanism in feature space. Our network architecture allows us to obtain an attention mask for each feature while existing CNN visualization methods provide only a common attention mask for all features. We apply the proposed multi-channel attention mechanism to multi-attribute recognition task. We can obtain different attention mask for each feature and for each attribute. Those analyses give us deeper insight into the feature space of CNNs. The experimental results for the benchmark dataset show that the proposed method gives high interpretability to humans while accurately grasping the attributes of the data.

cs.CV↗

arXiv2019

The brightest UV-selected galaxies in protoclusters at $z\sim4$: Ancestors of Brightest Cluster Galaxies?

We present the results of a survey of the brightest UV-selected galaxies in protoclusters. These proto-brightest cluster galaxy (proto-BCG) candidates are drawn from 179 overdense regions of $g$-dropout galaxies at $z\sim4$ from the Hyper Suprime-Cam Subaru Strategic Program identified previously as good protocluster candidates. This study is the first to extend the systematic study of the progenitors of BCGs from $z\sim2$ to $z\sim4$. We carefully remove possible contaminants from foreground galaxies and, for each structure, we select the brightest galaxy that is at least 1 mag brighter than the fifth brightest galaxy. We select 63 proto-BCG candidates and compare their properties with those of galaxies in the field and those of other galaxies in overdense structures. The proto-BCG candidates and their surrounding galaxies have different rest-UV color $(i - z)$ distributions to field galaxies and other galaxies in protoclusters that do not host proto-BCGs. In addition, galaxies surrounding proto-BCGs are brighter than those in protoclusters without proto-BCGs. The image stacking analysis reveals that the average effective radius of proto-BCGs is $\sim28\%$ larger than that of field galaxies. The $i-z$ color differences suggest that proto-BCGs and their surrounding galaxies are dustier than other galaxies at $z\sim4$. These results suggest that specific environmental effects or assembly biasses have already emerged in some protoclusters as early as $z \sim 4$, and we suggest that proto-BCGs have different star formation histories than other galaxies in the same epoch.

astro-ph.GA↗

arXiv2019

Improving Transparency of Deep Neural Inference Process

Deep learning techniques are rapidly advanced recently, and becoming a necessity component for widespread systems. However, the inference process of deep learning is black-box, and not very suitable to safety-critical systems which must exhibit high transparency. In this paper, to address this black-box limitation, we develop a simple analysis method which consists of 1) structural feature analysis: lists of the features contributing to inference process, 2) linguistic feature analysis: lists of the natural language labels describing the visual attributes for each feature contributing to inference process, and 3) consistency analysis: measuring consistency among input data, inference (label), and the result of our structural and linguistic feature analysis. Our analysis is simplified to reflect the actual inference process for high transparency, whereas it does not include any additional black-box mechanisms such as LSTM for highly human readable results. We conduct experiments and discuss the results of our analysis qualitatively and quantitatively, and come to believe that our work improves the transparency of neural networks. Evaluated through 12,800 human tasks, 75% workers answer that input data and result of our feature analysis are consistent, and 70% workers answer that inference (label) and result of our feature analysis are consistent. In addition to the evaluation of the proposed analysis, we find that our analysis also provide suggestions, or possible next actions such as expanding neural network complexity or collecting training data to improve a neural network.

cs.LG↗

arXiv2019

First release of high-redshift superluminous supernovae from the Subaru HIgh-Z sUpernova CAmpaign (SHIZUCA). II. Spectroscopic properties

We present Keck spectroscopic observations of three probable high redshift superluminous supernovae (SLSNe) from the Subaru HIgh-Z sUpernova CAmpaign (SHIZUCA), confirming redshifts of 1.851, 1.965 and 2.399. The host galaxies were selected for transient monitoring from multi-band photometric redshifts. The supernovae are detected during their rise, and the classically scheduled spectra are collected near maximum light. The restframe far-ultraviolet (FUV;$\sim$1000A-2500A) spectra include a significant host galaxy flux contribution and we compare our host galaxy subtracted spectra to UV-luminous SNe from the literature. While the signal-to-noise ratios of the spectra presented here are sufficient for redshift confirmation, supernova spectroscopic type confirmation remains inconclusive. The success of the first SHIZUCA Keck spectroscopic follow-up program demonstrates that campaigns such as SHIZUCA are capable of identifying high redshift SLSNe with sufficient accuracy, speed and depth for rapid, well-cadenced and informative follow-up.

astro-ph.HE↗

arXiv2019

First release of high-redshift superluminous supernovae from the Subaru HIgh-Z sUpernova CAmpaign (SHIZUCA). I. Photometric properties

We report our first discoveries of high-redshift supernovae from the Subaru HIgh-Z sUpernova CAmpaign (SHIZUCA), the transient survey using Subaru/Hyper Suprime-Cam. We report the discovery of three supernovae at the spectroscopically-confirmed redshifts of 2.399 (HSC16adga), 1.965 (HSC17auzg), and 1.851 (HSC17dbpf), and two supernova candidates with the host-galaxy photometric redshifts of 3.2 (HSC16apuo) and 4.2 (HSC17dsid), respectively. In this paper, we present their photometric properties and the spectroscopic properties of the confirmed high-redshift supernovae are presented in the accompanying paper (Curtin et al. 2019). The supernovae with the confirmed redshifts of z ~ 2 have the rest ultraviolet peak magnitudes close to -21 mag and they are likely superluminous supernovae. The discovery of three supernovae at z ~ 2 roughly corresponds to the approximate event rate of ~ 900 +/- 520 Gpc-3 yr-1 with Poisson error, which is consistent with the total superluminous supernova rate estimated by extrapolating the local rate based on the cosmic star-formation history. Adding unconfirmed superluminous supernova candidates would increase the event rate. Our superluminous supernova candidates at the redshifts of around 3 and 4 indicate the approximate superluminous supernova rates of ~ 400 +/- 400 Gpc-3 yr-1 (z ~ 3) and ~ 500 +/- 500 Gpc-3 yr-1 (z ~ 4) with Poisson errors. Our initial results demonstrate the outstanding capability of Hyper Suprime-Cam to discover high-redshift supernovae.

astro-ph.HE↗

arXiv2019

Automatic Labeled LiDAR Data Generation based on Precise Human Model

Following improvements in deep neural networks, state-of-the-art networks have been proposed for human recognition using point clouds captured by LiDAR. However, the performance of these networks strongly depends on the training data. An issue with collecting training data is labeling. Labeling by humans is necessary to obtain the ground truth label; however, labeling requires huge costs. Therefore, we propose an automatic labeled data generation pipeline, for which we can change any parameters or data generation environments. Our approach uses a human model named Dhaiba and a background of Miraikan and consequently generated realistic artificial data. We present 500k+ data generated by the proposed pipeline. This paper also describes the specification of the pipeline and data details with evaluations of various approaches.

cs.CV↗

arXiv2019

Effect of lensing magnification on type Ia supernova cosmology

Effect of gravitational magnification on the measurement of distance modulus of type Ia supernovae is presented. We investigate a correlation between magnification and Hubble residual to explore how the magnification affects the estimation of cosmological parameters. We estimate magnification of type Ia supernovae in two distinct methods: one is based on convergence mass reconstruction under the weak lensing limit and the other is based on the direct measurement from galaxies distribution. Both magnification measurements are measured from Subaru Hyper Suprime-Cam survey catalogue. For both measurements, we find no significant correlation between Hubble residual and magnification. Furthermore, we correct for the apparent supernovae fluxes obtained by Supernova Legacy Survey 3-year sample using direct measurement of the magnification. We find $Ω_{\rm m0} = 0.287 ^{+0.104} _{-0.085}$ and $w = -1.161 ^{+0.595}_{-0.358}$ for supernovae samples corrected for lensing magnification when we use photometric redshift catalogue of Mizuki, while $Ω_{\rm m0} = 0.253 ^{+0.113} _{-0.087}$ and $w = -1.078 ^{+0.498} _{-0.297}$ for DEmP photo-z catalogue. Therefore, we conclude that the effect of magnification on the supernova cosmology is negligibly small for the current surveys; however, it has to be considered for the future supernova survey like LSST.

astro-ph.CO↗