Search

With the increasing computing power of edge devices, Federated Learning (FL) emerges to enable model training without privacy concerns. The majority of existing studies assume the data are fully labeled on the client side. In practice, however, the amount of labeled data is often limited. Recently, federated semi-supervised learning (FSSL) is explored as a way to effectively utilize unlabeled data during training. In this work, we propose ProtoFSSL, a novel FSSL approach based on prototypical networks. In ProtoFSSL, clients share knowledge with each other via lightweight prototypes, which prevents the local models from diverging. For computing loss on unlabeled data, each client creates accurate pseudo-labels based on shared prototypes. Jointly with labeled data, the pseudo-labels provide training signals for local prototypes. Compared to a FSSL approach based on weight sharing, the prototype-based inter-client knowledge sharing significantly reduces both communication and computation costs, enabling more frequent knowledge sharing between more clients for better accuracy. In multiple datasets, ProtoFSSL results in higher accuracy compared to the recent FSSL methods with and without knowledge sharing, such as FixMatch, FedRGD, and FedMatch. On SVHN dataset, ProtoFSSL performs comparably to fully supervised FL methods.

Context : AKARI is the first Japanese astronomical satellite dedicated to infrar ed astronomy. One of the main purposes of AKARI is the all-sky survey performed with six infrared bands between 9 and 200um during the period from 2006 May 6 to 2007 August 28. In this paper, we present the mid-infrared part (9um and 18um b ands) of the survey carried out with one of the on-board instruments, the Infrar ed Camera (IRC). Aims : We present unprecedented observational results of the 9 and 18um AKARI al l-sky survey and detail the operation and data processing leading to the point s ource detection and measurements. Methods : The raw data are processed to produce small images for every scan and point sources candidates, above the 5-sigma noise level per single scan, are der ived. The celestial coordinates and fluxes of the events are determined statisti cally and the reliability of their detections is secured through multiple detect ions of the same source within milli-seconds, hours, and months from each other. Results : The sky coverage is more than 90% for both bands. A total of 877,091 s ources (851,189 for 9um, 195,893 for 18um) are confirmed and included in the cur rent release of the point source catalogue. The detection limit for point source s is 50mJy and 90mJy for the 9um and 18um bands, respectively. The position accu racy is estimated to be better than 2". Uncertainties in the in-flight absolute flux calibration are estimated to be 3% for the 9um band and 4% for the 18um ban d. The coordinates and fluxes of detected sources in this survey are also compar ed with those of the IRAS survey and found to be statistically consistent., Comment: Accepted for publication in AandA AKARI special issue

The Infrared Camera (IRC) is one of the two instruments on board the AKARI satellite. In addition to deep imaging from 1.8-26.5um for the pointed observation mode of the AKARI, it has a spectroscopic capability in its spectral range. By replacing the imaging filters by transmission-type dispersers on the filter wheels, it provides low-resolution (lambda/d_lambda ~ 20-120) spectroscopy with slits or in a wide imaging field-of-view (approximately 10'X10'). The IRC spectroscopic mode is unique in space infrared missions in that it has the capability to perform sensitive wide-field spectroscopic surveys in the near- and mid-infrared wavelength ranges. This paper describes specifications of the IRC spectrograph and its in-orbit performance., Comment: 13 pages, 7 figures, accepted for publication on PASJ

We present the results of optical identifications for 257 mid-infrared sources detected with a deep 15um survey over approximately 80 arcmin^2 area in the AKARI performance verification field near the North Ecliptic Pole. The 15um fluxes of the sources range from 1 mJy down to 40 uJy, approximately a half of which are below 100 uJy. Optical counterparts were searched for within a 2-3 arcsec radius in both the BVRi'z' catalog generated by using the deep Subaru/Suprime-cam field which covers one-third of the performance verification field, and the g'r'i'z' catalog based on observations made with MegaCam at CFHT. We found B-R and R-z' colours of sources with successful optical identifications are systematically redder than that of the entire optical sample in the same field. Moreover, approximately 40% of the 15um sources show colours R-L15>5, which cannot be explained by the spectral energy distribution (SED) of normal quiescent spiral galaxies, but are consistent with SEDs of redshifted (z>1) starburst or ultraluminous infrared galaxies. This result indicates that the fraction of the ultraluminous infrared galaxies in our faint 15um sample is much larger than that in our brighter 15um sources, which is consistent with the evolving mid-infrared luminosity function derived by recent studies based on the Spitzer 24um deep surveys. Based on an SED fitting technique, the nature of the faint 15um sources is further discussed for a selected number of sources with available K_s-band data.

We report the first results of a near- and mid- infrared deep survey with the Infrared Camera (IRC) onboard AKARI in the performance verification phase. Simultaneous observations by the NIR, MIR-S and MIR-L channels of the IRC with effective integration times of 4529, 4908, and 4417 seconds at 3, 7, and 15 micron, covering 86.0, 70.3, and 77.3 arcmin^2 area, detected 955, 298 and 277 sources, respectively. The 5 sigma detection limits of the survey are 6.0, 31.5 and 71.2 micro Jy and the 50% completeness limit are 24.0, 47.5, and 88.1 micro Jy at 3, 7, and 15 micron, respectively. The observation is limited by source confusion at 3 micron. We have confirmed the turnover in the 15 micron differential source counts around 400 micro Jy, previously detected by surveys with the Infrared Space Observatory. The faint end of 15 micron raw source counts agree with the results from the deep surveys in the GOODS fields carried out with the Spitzer IRS peak up imager and the predictions of current galaxy evolution models. These results indicate that deep surveys with comprehensive wavelength coverage at mid-infrared wavelength are very important to investigate the evolution of infrared galaxies at high redshifts., Comment: 15 pages, 13 figures, and 5 tables. Accepted for publication in the AKARI special issue of the Publications of the Astronomical Society of Japan

We present the observations of the reflection nebulae IC4954 and IC4955 region with the Infrared Camera (IRC) and the Far-Infrared Surveyor (FIS) on board the infrared astronomical satellite AKARI during its performance verification phase. We obtained 7 band images from 7 to 160um with higher spatial resolution and higher sensitivities than previous observations. The mid-infrared color of the S9W (9um) and L18W (18um) bands shows a systematic variation around the exciting sources. The spatial variation in the mid-infrared color suggests that the star-formation in IC4954/4955 is progressing from south-west to north-east. The FIS data also clearly resolve two nebulae for the first time in the far-infrared. The FIS 4-band data from 65um to 160um allow us to correctly estimate the total infrared luminosity from the region, which is about one sixth of the energy emitted from the existing stellar sources. Five candidates for young stellar objects have been detected as point sources for the first time in the 11um image. They are located in the red S9W to L18W color regions, suggesting that current star-formation has been triggered by previous star-formation activities. A wide area map of the size of about 1 x 1 (deg^2) around the IC4954/4955 region was created from the AKARI mid-infrared all-sky survey data. Together with the HI 21cm data, it suggests a large hollow structure of a degree scale, on whose edge the IC4954/4955 region has been created, indicating star formation over three generations in largely different spatial scales., Comment: 23 pages, 7 figures, accepted for publication in PASJ AKARI special issue

We report the detection of an H-alpha emission line in the low resolution spectrum of a quasar, RX J1759.4+6638, at a redshift of 4.3 with the InfraRed Camera (IRC) onboard the AKARI. This is the first spectroscopic detection of an H-alpha emission line in a quasar beyond z=4. The overall spectral energy distribution (SED) of RX J1759.4+6638 in the near- and mid-infrared wavelengths agrees with a median SED of the nearby quasars and the flux ratio of F(Ly-alpha)/F(H-alpha) is consistent with those of previous reports for lower-redshift quasars., Comment: 9pages, 3 figures, Publications of the Astronomical Society of Japan, in press

We report preliminary results of AKARI observations of two globular clusters, NGC104 and NGC362. Imaging data covering areas of about 10x10 arcmin^2 centered on the two clusters have been obtained with InfraRed Camera (IRC) at 2.4, 3.2, 4.1, 7.0, 9.0, 11.0, 15.0, 18.0 and 24.0 mu. We used F11/F2 and F24/F7 flux ratios as diagnostics of circumstellar dust emission. Dust emissions are mainly detected from variable stars obviously on the asymptotic giant branch, but some variable stars that reside below the tip of the first-ascending giant branch also show dust emissions. We found eight red sources with F24/F7 ratio greater than unity in NGC362. Six out of the eight have no 2MASS counterparts. However, we found no such source in NGC104., Comment: Accepted for publication in PASJ

We present the properties of 11 $\mu$m selected sources detected in the early data of the North Ecliptic Pole Deep (NEP-Deep) Survey of AKARI. The data set covers 6 wavelength bands from 2.5 to 11 $\mu$m, with the exposure time of 10 ~ 20 minutes. This field lies within the CFHT survey with four filter bands ($g', r', i',z'), enabling us to establish nearly continuous spectral energy distributions (SEDs) for wavelengths ranging from 0.4 to 11 $\mu$m. The main sample studied here consists of 71 sources whose 11 $\mu$m AB magnitudes are equal to or brighter than 18.5 (251 $\mu$Jy), which is complete to more than 90%. The 11 $\mu$m band has an advantage of sampling star forming galaxies with low to medium redshifts since the prominent PAH feature shifts into this band. As expected, we find that the majority (~68%) of 11 $\mu$m bright sources are star forming galaxies at 0.2 < z < 0.7 with $L_{IR} ~ 10^{10}$ -- $10^{12} L_{\odot}$ based on the detailed modelling of SEDs. We also find four AGNs lying at various redshifts in the main sample. In addition, we discuss a few sources which have non-typical SEDs of the main sample, including a brown dwarf candidate, a steep power-law source, flat spectrum object, and an early-type galaxy at moderate redshift., Comment: 22 pages, 10 figures

AKARI (formerly ASTRO-F) is an infrared space telescope designed for an all-sky survey at 10-180 (mu)m, and deep pointed surveys of selected areas at 2-180 (mu)m. The deep pointed surveys with AKARI will significantly advance our understanding of galaxy evolution, the structure formation of the Universe, the nature of the buried AGNs, and the cosmic infrared background. Here we describe the important characteristics of the AKARI mission: the orbit, and the attitude control system, and investigate the optimum survey area based on the updated pre-flight sensitivities of AKARI, taking into account the cirrus confusion noise as well as the surface density of bright stars. The North Ecliptic Pole (NEP) is concluded to be the best area for 2-26 (mu)m deep surveys, while the low-cirrus noise regions around the South Ecliptic Pole (SEP) are worth considering for 50-180 (mu)m pointed surveys to high sensitivities limited by the galaxy confusion noise. Current observational plans of these pointed surveys are described in detail. Comparing these surveys with the deep surveys with the Spitzer Space Telescope, the AKARI deep surveys are particularly unique in respect of their continuous wavelength coverage over the 2-26 (mu)m range in broad-band deep imaging, and their slitless spectroscopy mode over the same wavelength range., Comment: accepted for publication in Publications of the Astronomical Society of Japan Vol.58(4) (2006), 39 pages, 16 Postscript figures, uses pasj00.cls