Filename
stringlengths 22
64
| Paragraph
stringlengths 8
5.57k
|
|---|---|
Processed_Hazard_function_models_to_estimate_mortality_rates.txt | year and habitat to obtain a dataset representative of the entire catch in this fishery. |
Processed_Hazard_function_models_to_estimate_mortality_rates.txt | Figure 1: Comparison between simulated natural mortality (x-axis) and estimated (y-axis) using a random sample of the matrix of catch (strategy 1); a random sample from each year separately (strategy 2) and the same sample weighted by yearly total catch (strategy 2 - weighted samples). Each panel correspond to an increasing number of samples per year varying from 125 to 2000. |
Processed_Hazard_function_models_to_estimate_mortality_rates.txt | Figure 2: Comparison between simulated catchability (x-axis) and estimated (y-axis) using a random sample of the matrix of catch (strategy 1); a random sample from each year separately (strategy 2) and the same sample weighted by yearly total catch (strategy 2 - weighted samples). Each panel correspond to an increasing number of samples per year varying from 125 to 2000. |
Processed_Hazard_function_models_to_estimate_mortality_rates.txt | Figure 3: Difference between the negative log-likelihood (−log(L)) from survival analysis (SA) and multinomial (ML) as a function of the number of sample per year. Each panel represents a particular sampling strategy. |
Processed_Hazard_function_models_to_estimate_mortality_rates.txt | it used the simulated proportions of individuals at age in place of the probabilities to compute the likelihood. |
Processed_Hazard_function_models_to_estimate_mortality_rates.txt | to identify the gear selectivity that best fit the data, just not an automatic one. |
Processed_Hazard_function_models_to_estimate_mortality_rates.txt | Table 2: Distribution of yearly samples (in rows) of sea mullet into age-groups of width 1 year (in columns) in the estuaries and ocean habitats; catch in tonnes and effort in number of days. |
Processed_Hazard_function_models_to_estimate_mortality_rates.txt | Table 3: Comparison between the log-likelihood value obtained for several hazard function models of the mullet data using different numbers of parameters (p). The models were ordered by increasing values of Akaike Information Criteria (AIC =−2log(L) + 2p) from top to bottom. |
Processed_Hazard_function_models_to_estimate_mortality_rates.txt | Table 5: Convention used to associate each element of the catch at age matrix (Ci,j ) with particular cohort referred to as with the number given in this table. |
Processed_Hazard_function_models_to_estimate_mortality_rates.txt | D.R. Cox and D. Oakes. Analysis of survival data. Chapman and Hall Ltd, London, 1984. |
Processed_Hazard_function_models_to_estimate_mortality_rates.txt | G.L. Kesteven. Further results of tagging sea mullet, Mugil cephalus Linnaeus, on the eastern Australian coast. |
Processed_Hazard_function_models_to_estimate_mortality_rates.txt | T. J. Quinn and R. B. Deriso. Quantitative Fish Dynamics. Oxford University Press, New York, 1999. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | The lack of computational power within an organization for analyzing scientific data, and the distribution of knowledge (by scientists) and technologies (advanced scientific devices) are two major problems commonly observed in scientific disciplines. One such scientific discipline is brain science. The analysis of brain activity data gathered from the MEG (Magnetoencephalography) instrument is an important research topic in medical science since it helps doctors in identifying symptoms of diseases. The data needs to be analyzed exhaustively to efficiently diagnose and analyze brain functions and requires access to large-scale computational resources. The potential platform for solving such resource intensive applications is the Grid. This paper describes a MEG data analysis system developed by us, leveraging Grid technologies, primarily Nimrod-G, Gridbus, and Globus. This paper explains the application of economy-based grid scheduling algorithms to the problem domain for on-demand processing of analysis jobs. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | The emergence of high speed networks has made it possible to share geographically distributed resources such as supercomputers, storage systems, databases and scientific instruments in order to gather, process and transfer data smoothly across different administrative domains. Aggregations of such distributed resources, called computational grids[1], provide computing power that has made it possible to solve large scale problems in science, engineering and commerce. Biological sciences have several computational and data intensive problems which have not been tackled satisfactorily for the want of adequate computing and storage resources. The analysis of data gathered by monitoring activity in the brain is one such application. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | Brain activity is measured by the Magnetoencephalography (MEG) instrument which measures the magnetic fields generated by the electrical activity in the brain. This method is more accurate than others such as Electroencephalography (EEG) and Electrocorticography (ECoG)[2]. Another advantage of MEG is that it is non-invasive. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | track a signal that is within the spectrum. This enables mapping of brain activity to patient behaviour and is useful to diagnose several diseases. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | However, there are certain problems with MEG. Due to the high cost of equipment and the required shielding against stray magnetic fields, there are only limited numbers of MEG instruments around the world. Plus, the instrument generates a huge amount of data, all of which are not analysed due to of lack of computing resources. For example, a 64-sensor MEG instrument would produce 0.9GB of data over a period of an hour. Such a task generates 7,257,600 analysis jobs and would take 102 days on a commodity computer with a PentiumIII/500MHz processor and 256MB of memory. Proper recovery of the patient depends on the results being available as soon as possible so that the doctor can start the drug regimen in the shortest possible time. While it is possible to use a supercomputer to analyse the brain data, every medical facility cannot be expected to have access to such a machine. Also, access to a much larger pool of resources and consequently, larger amounts of computational power is possible if the application can be deployed on the Grid. Furthermore, there is also the advantage that the medical data can be shared easily among the partnering doctors. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | The rest of the paper is organized as follows: Section 2 describes the flow diagram for the analysis. Section 3 describes the architecture for the execution. Section 4 explains the process of grid-enabling this application. The scheduling experiments and the results of those are detailed in Section 5. Section 6 presents related work and the final section summarizes the paper along with suggestions for future works. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | The NeuroGrid project aims to convert the existing brain activity analysis application into a parameter sweep application for executing jobs which perform wavelet cross-correlation analysis for each pair of sensors in parallel on distributed resources. The jobs are both computationally and data intensive. Also they are independent of each other thus making them perfect for being executed on the Grid. The flow diagram for distributed analysis of brain activity is shown in Fig. 1. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | time. The data and the analysis code are dispatched to the remote node and the results collected (steps 4 and 5). |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | The architecture followed in this project is shown in Fig. 2. It consists of brain activity analysis application, paramterisation tools (Nimrod-G parameter specification language), resource broker (Nimrod-G with Gridbus scheduler), grid market directory (Gridbus GMD), and low-level grid middleware (Globus[4]). The resources are Grid-enabled using Globus software deployed on them. The application has been Grid enabled by composing it as a parameter sweep application using the Nimrod-G parameter specification language. The GMD has been used as a register for publication of resource providers and services. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | The analysis code was developed by the Cybermedia Centre, Osaka University, Japan. The raw data obtained from the sensors in the MEG instrument is analysed in two phases as shown in the Fig. 3. In the first phase, the raw data from the brain goes through wavelet transform operation. This phase gives the time-frequency data of the output. In the next phase, cross-correlation analysis is performed for each pair of wavelet transforms. This output displays the similarity between a pair of brain data in every frequency spectrum. By focusing on a specific frequency spectrum, the sensor that first detected the signal with that focusing spectrum can be localized. This means that the path of a signal can be found as it travels through the brain. The code was written in C and is highly portable. We have been able to compile it for different platforms such as PC/Linux, Sun/Solaris, SGI/Irix, Alpha, etc. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | The Gridbus scheduler developed as a plugin scheduler for Nimrod-G has been used instead of the default Nimrod-G scheduler since it has been designed to utilise the Grid Market Directory (GMD) [6].The default Nimrod-G scheduler calculates the processing cost based on the CPU time that is used to execute a job on a remote node. To support the notion of application services and pricing based on AO (Application Operation) instead of vanilla CPU service, the GMD already allows the GSPs (Grid Service Providers) to publish application services along with their AO service price. Hence, the Gridbus scheduler has been enabled to utilise the GMD services and perform resource allocation based on AO cost model. In this model, the user is charged a price for execution of each job on the resource. Thus, the resource owner may offer the application as a service and charge a fixed price for executing it. By periodically updating the resource information from the GMD, the Gridbus scheduler ensures that the scheduling is done based on the most recent cost-price for the resources. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | The Gridbus Scheduler implements three algorithms: cost minimization, time minimization and cost-time optimization. All three algorithms are constrained by two parameters: the deadline by which the experiment is required to complete and the budget that the user has. Time minimization tries to execute the project within the shortest time while keeping within the budget. Cost minimization tries to complete the execution with the least cost while keeping to the deadline. Cost-time optimization gives jobs to the cheapest servers but performs time optimization among those. Cost-time optimization was only simulated until recently and it has been implemented for the first time in the Gridbus Scheduler. More information about these scheduling algorithms can be found in [7, 8]. The scheduler uses the past performance of each machine to forecast the job completion rate of that machine. Also, it averages the job completion rate so that any spikes or troughs in performance are smoothed out. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | The Grid Market Directory (GMD)[6], developed by the Gridbus project in the University of Melbourne, allows service providers to publish the services which they provide along with the costs associated with them which the consumers are able to browse and locate a service which meets their requirements. GMD is built over standard Web service technologies such as SOAP (Simple Object Access Protocol) and XML(eXtensible Markup Language) therefore it can be queried by other client programs. To provide with an additional layer of transparency, a client API (Application Programming Interface) has been provided that could be used by programs to query the GMD without the developers having to concern themselves with details of SOAP. The GMD infrastructure was used to maintain a registry of participants along with the details of their contributed resources as part of the Global Grid Testbed collaboration. The Gridbus scheduler interacts with the GMD to discover the testbed resources and their high-level attributes such as access price. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | The existing analysis code was composed as a task-farming, parameter sweep application for execution on the Grid using the Nimrod-G parameter specification language[5]. The Nimrod-G farming engine and dispatcher along with Gridbus scheduler is used for deploying and processing it on Global Grids. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | The brain activity analysis software suite, developed at the Cybermedia Centre at Osaka University, consists of two separate programs. The first, raw2wavelet, is a program that performs wavelet transform over the raw data and the second, wavelet2cross, is a program that does the cross-correlation between the wavelet transforms. A job consists of executing these two programs in sequence for a pair of unique sensors and for a particular offset. For 64 sensors and an offset ranging over 0 to 29,750, this would generate (64*63)/2 * 29750 jobs. Although this application contains numerous jobs, each individual job is fine-grained in nature--in certain data scenarios it can be less than a minute. The overhead associated with initiating each task on a separate node and collecting its results after it finished execution would have radically decreased the efficiency of distributed execution. So, coarse grained jobs had to be created by grouping the fine-grained jobs so that the computation to I/O ratio would work in favour of distributed analysis. For this purpose, a meta-program (metameg) was written that would perform pairwise analysis for all sensors for a particular range of offsets. The entire temporal shift region was divided into offset ranges. A job would consist of performing analysis over a temporal shift range for all the sensors. The pseudo-code for this meta job program, called metameg, is listed in Fig. 4. |
Processed_Economic_and_On_Demand_Brain_Activity_Analysis_on_.txt | The executables of the three programs that represented the analysis code (metameg, raw2wavelet, and wavelet2cross implemented in C language) were small in size and hence easily transferable at runtime if the code is not deployed on remote resources. Each meta job requires the full raw data captured by all (64) sensors. However, transferring this raw data, over 24MB, to each remote node during the execution any job would constitute a significant I/O load over wide-area networks. Unless an analysis is carried out on a MEG data with few sensors generating small amount of data, it is preferable to pre-stage the entire raw dataset at each node. We also had to change the meta program (metameg) and the raw data to wavelet conversion (raw2wavelet) program to enable access to data on a path on the remote node. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | ABSTRACT A deep survey of the European Large Area ISO Survey (ELAIS) field in the southern celestial hemisphere (hereinafter S1) has been carried out with the Australia Telescope Compact Array (ATCA) at 1.4 GHz. The S1 region, covering about 4 square degrees, has been also surveyed in the mid- and far-infrared (5-200 µm) with the Infrared Space Observatory (ISO). The radio survey provides uniform coverage of the entire S1 region, with a sensitivity (5σ) of 0.4 mJy over the whole area and 0.2 mJy in the center. To this sensitivity, virtually all the radio counterparts of the far-infrared extragalactic ISO sources should be detected. This paper presents a radio sample–complete at the 5σ level–consisting of 581 sources detected at 1.4 GHz. Of these radio sources, 349 have peak flux density in the range 0.2-1 mJy, forming a new homogeneous sample of sub-mJy radio sources. Due to its size, depth and multi-waveband coverage, the sample will allow us to study in greater detail the sub-mJy radio source population. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | Key words: cosmology: observations – radio continuum: galaxies – infrared: galaxies – surveys – galaxies: star-burst. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | higher z, and possibly unveiling new classes of objects or unexpected phenomena. Since dust plays an important role in most of these goals, and since many of the thousands of galaxies detected in the ELAIS survey will be at high red- shift and probably obscured in the optical bands, radio ob- servations will play a crucial role in assessing the reliability of the ELAIS source list and in facilitating source identifica- tion in faint or empty optical fields. Additionally, radio data will be important in the optical identification phase because the spatial resolution of ISO is insufficient to identify un- ambiguously many of the faintest ISO sources. Even at 15 µm the survey resolution is 10 arcsec, while at 90 µm it ∼ is almost 1 arcmin, so there are multiple optical candidates within each error box. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | The three northern ELAIS fields (N1, N2 and N3) have been surveyed in the radio band (at 1.4 GHz) with the Very Large Array (VLA) down to an average flux density limit of 0.25 mJy (5σ) over a total area of 4.22 deg2. Details of the VLA observations and a description of the catalogue containing 867 radio sources can be found in Ciliegi et al. (1998). |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | In this paper we describe radio observations of the ELAIS field located in the southern celestial hemisphere, S1. This field, centered at α(2000) = 00h 34m 44.4s, δ(2000) = -43◦ 28′ 12′′, covers an area of the sky of about 2◦ 2◦. We observed the whole area at 1.4 GHz with the Australia Telescope Compact Array (ATCA) in the 6-km configura- tion (maximum baseline length), with a resolution of 8 15 arcsec. We obtained uniform radio coverage over the whole S1 region, with sufficient sensitivity (1σ 80 µJy) to detect virtually all radio counterparts of the ISO galaxies. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | By combining these radio observations with the avail- able ISO data, we will investigate the radio/far-infrared cor- relation (Helou, Soifer & Rowan-Robinson 1985) in star- forming galaxies to a flux density significantly deeper than that reached by IRAS. When spectral informations become available for the optical counterparts of the ISO and ATCA sources in the field, we will investigate the trivariate IR- radio-optical luminosity function and its evolution. This study will be of fundamental importance for interpreting the source counts at different wave-bands and for elucidat- ing contributions from different classes of objects. Moreover, we will be able to determine the influence of obscuration on the inference of the star-formation rate, and to constrain the star formation history of the Universe to z 1.0–1.5 with two complementary samples, selected in the infrared and ra- dio, within the same volume of the Universe (see Cram et al. 1998; Oliver, Gruppioni & Serjeant 1998). |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | hint strongly at an increasing fraction of early-type galaxies (Gruppioni, Mignoli & Zamorani 1998). This suggestion, to- gether with apparently discordant results from optical iden- tifications of even fainter µJy radio samples ( Windhorst et al. 1995; Hammer et al. 1995; Richards et al. 1998) com- plicate the picture. The ATCA survey in the ELAIS field S1, together with the VLA surveys of the northern fields N1, N2 and N3 (Ciliegi et al. 1998) will shed new light on the sub-mJy population and its connection to the faint blue and infrared star-burst galaxies. The sensitivity, areal cov- erage and multi-wavelength character (optical, near-, mid- and far-infrared and radio) of the ELAIS project will en- sure that it makes an important contribution to all of these problems. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | In section 2 of the paper we describe the radio obser- vations. Section 3 discusses the data reduction strategy and section 4 presents the source catalogue. In section 5 we dis- cuss the radio source counts and present our conclusions in section 6. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | By choosing an observing frequency of 1.4 GHz, we ensure a large instantaneous field of view (33 arcmin primary beam) and hence shorten the observing time necessary to cover the entire S1 area. The sensitivity is also higher for the same integration time with respect to higher available frequen- cies (2.4, 4.86 and 8.44 GHz). Moreover, a frequency of 1.4 GHz reveals well any recent star-formation activity, seen at this frequency as synchrotron radiation excited by supernova remnants (Condon 1992). |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | where Ssys (in Jy) is the system sensibility (350 Jy at 1.4 GHz), t is the integration time (in minutes), B is the band- width (in MHz) and N is the number of baselines. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | The ATCA is an east-west array which uses the earth’s ro- tation to sample visibilities over elliptical loci in the spa- tial frequency (u, v) plane. Its control system allows imag- ing over a field or a source whose extent is larger than the primary beam by cycling through a grid of pointings on the sky, recording the visibilities for each pointing intermit- tently. This mosaic observing mode is an efficient way to obtain uniform quality imaging over a large field, and has been used to construct our survey. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | Figure 1. The sky position and orientation of the ISO S1 survey region. The outer rectangle is 2◦ × 2◦. The filled triangles show the ISO-CAM 15 µm pointings, while the filled squares show the ISO-PHOT 90 µm pointings. The circles (drawn with a radius of 20 arcmin) show the ATCA mosaic pointings. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | requires greater computing resources and is less suited to images dominated by compact sources. Linear combination mosaicing takes a pixel-by-pixel weighted mean of the sin- gle pointings, where the weights are determined from the primary beam response and the noise level in that pixel (Sault & Killeen 1995). The value of each pixel in the fi- nal mosaic depends on the pointing configuration and, in particular, on the grid spacing. For our survey, an optimal trade-off between uniformity of sensitivity and efficiency of telescope usage was obtained by adopting a rectangular grid 7 fields, with a regular spacing between fields equal of 7 to θF W HP /√2 20 arcmin (where θF W HP is the full width at half power of the primary beam, 33 arcmin at 1.4 GHz). Figure 1 shows the sky position and orientation of the S1 region. Circles drawn with a radius of 20 arcmin show the 49 ATCA pointings. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | To optimize (u, v) coverage we organised the observa- tions by dividing the 49 fields in four blocks of 13 fields each and observing each block for 12h + 3h (one night plus one quarter of the last night). During this time the 13 fields were observed with dwell times of 20 seconds each, adding 3 min- utes for observing the secondary calibrator every hour (i.e. every 15 cycles of 13 fields). The field corresponding to the center of the S1 area was visited in each block, to obtain deeper sensitivity in the central 20 arcmin of the ELAIS field. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | The primary flux density calibrator was PKS B1934- 638, whose flux densities at different frequencies are incor- porated directly in the calibration software. The primary calibrator was observed for 10 min at the beginning of each 12h observing run. The source PKS B0022-423 was used as a phase and secondary amplitude calibrator. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | Before observing we checked for the presence of strong radio sources in the field, which could eventually compro- mise the sensitivity of the survey. Although the ELAIS S1 field was not selected with radio observations in mind, for- tunately there are only a few bright radio sources within it, so that the target noise level was obtained without seri- ous problems of confusion. The Parkes-MIT-NRAO (PMN) Southern Survey (Griffith & Wright 1993) reveals one bright radio source (PMN J0042-4413) just outside the field, with a flux density of 1.2 Jy at 5 GHz. There are also one 0.4 Jy and one 0.2 Jy PMN source in the field, and several with lower flux densities. We have observed the whole S1 area, expecting that not more than a small part of the mosaic might be degraded by the sidelobes of strong sources in or just outside the field. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | We used both receivers at 1.4 GHz, each with a band- pass of 128 MHz. Using the ATCA in this mode makes it hard to avoid strong interference in some channels for at least part of the observation, but the doubled bandwidth improves the sensitivity by a factor of √2. The optimal po- sitioning of the bands was obtained by centering one fre- quency at 1.344 GHz (1.280-1.408 GHz) and the other at 1.472 GHz (1.408-1.536 GHz). |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | The data were analyzed with the software package miriad (Multi-channel Image Reconstruction, Image Anal- ysis and Display), which is standard software for radio- interferometric data reduction specifically adapted to the ATCA. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | Figure 2. Grey scale image of the total 1.4 GHz mosaic in the ELAIS S1 field. The square region corresponds to the area observed by ISO. |
Processed_A_1.4_GHz_Survey_of_the_Southern_ELAIS_Region.txt | Before calibrating the data, we flagged all baselines and fringe visibilities affected by any problems encountered dur- ing the observation (such as antenna slewing intervals, inter- ference, etc) using the tasks blflag, tvflag and uvflag. Primary calibrator data were flagged before calibrating, and the secondary calibrator after application of the band-pass and instrumental polarization calibrations obtained from the primary calibrator. Finally, the fully calibrated single field data were flagged. Each bandpass (1.344 and 1.472 GHz) was calibrated separately, according to the standard calibration technique (Sault and Killeen 1995). |
Processed_Physical_Conditions_of_the_Gas_in_an_ALMA_[CII]-id.txt | We present 12CO(2–1) observations of the submillimetre galaxy ALESS65.1 per- formed with the Australia Telescope Compact Array at 42.3 GHz. A previous ALMA study of submillimetre galaxies in the Extended Chandra Deep Field South detected [CII] 157.74 µm emission from this galaxy at a redshift of z = 4.44. No 12CO(2–1) emis- sion was detected but we derive a firm upper limit to the cold gas mass in ALESS65.1 of Mgas < 1.7 × 1010 M(cid:12). The estimated gas depletion timescale is < 50 Myr, which is similar to other high redshift SMGs, and consistent with z > 4 SMGs being the likely progenitors of massive red-and-dead galaxies at z > 2. The ratio of the [CII], 12CO and far-infrared luminosities implies a strong far-ultraviolet field of G0 (cid:38) 103, as seen in Galactic star forming regions or local ULIRGs. The observed L[CII]/LFIR = 2.3 × 10−3 is high compared to local ULIRGs and, combined with L[CII]/LCO (cid:38) 2700, it is con- sistent with ALESS65.1 either having an extended (several kpc) [CII] emitting region or lower than solar metallicity. |
Processed_Physical_Conditions_of_the_Gas_in_an_ALMA_[CII]-id.txt | under-represented in SMG redshift surveys, mostly because they lie below the sensitivity limits of the radio interferome- ter surveys used to identify SMGs. However, in the last few years, over a dozen of these sources have been reported (Ca- pak et al. 2008, 2011; Daddi et al. 2009a,b; Coppin et al. 2009; Knudsen et al. 2010; Carilli et al. 2010, 2011; Riechers et al. 2010; Smolˇci´c et al. 2011; Cox et al. 2011; Combes et al. 2012; Walter et al. 2012), and recent ALMA redshift surveys have doubled these numbers (Weiss et al. submitted; Vieira et al. submitted). Detailed studies of the star forma- tion and gas content in these z > 4 SMGs are still rare, but are needed to provide a unique insight into the earliest phases of the growth of massive elliptical galaxies. |
Processed_Physical_Conditions_of_the_Gas_in_an_ALMA_[CII]-id.txt | The most accessible tracer of cold gas in galaxies is 12CO, which has been observed extensively in SMGs (e.g. Greve et al. 2005; Tacconi et al. 2006; Bothwell et al. 2013). Detections of 12CO in z > 4 SMGs have shown they are gas rich systems with sufficient reservoirs (Mgas > 1010 M(cid:12)) to sustain the extreme star formation rates of ∼ 1000 M(cid:12) yr−1 for only short time scales (10s of Myrs) (Coppin et al. 2010; Riechers et al. 2010), unless the gas is replenished. This is consistent with high redshift SMGs being the progenitors of the luminous red galaxies seen at z > 2. |
Processed_Physical_Conditions_of_the_Gas_in_an_ALMA_[CII]-id.txt | The 2P3/2 – 2P1/2 fine structure line of singly ionised carbon at 157.74 µm (here-after [CII]) has emerged as a powerful alternative line for studying the ISM in high red- shift sources. It can represent up to 1% of the bolomet- ric luminosity of star forming galaxies (e.g. Crawford et al. 1985; Stacey et al. 1991). This line emission arises mainly from the photodissociation regions (PDRs) that form at the edges of molecular clouds illuminated by the UV photons of young-massive stars, but a significant contribution can also come from HII regions and the more diffuse warm in- terstellar medium (Madden et al. 1993; Heiles 1994). The [CII] line therefore provides an important probe of the gas content and star formation processes in a galaxy. |
Processed_Physical_Conditions_of_the_Gas_in_an_ALMA_[CII]-id.txt | A recent ALMA Cycle 0 study of 126 submillimetre sources located in the LABOCA Extended Chandra Deep Field South (LESS, Weiß et al. 2009; Karim et al. 2012; Hodge et al. 2013 submitted) resulted in the serendipitous identification of [CII] line emission from two SMGs (Swin- bank et al. 2012, hereafter S12). The average [CII[/far- infrared luminosity ratio of these two SMGs is ∼ 0.0012 ± 0.0004, roughly ten times higher than that observed in local ultraluminous infrared galaxies, which has been in- terpreted as evidence that their gas reservoirs are more extended (S12). The large extent of SMGs is supported by other observations, such as extended radio morpholo- gies (e.g. Chapman et al. 2004; Biggs & Ivison 2008), ex- tended H-α morphologies (e.g. Swinbank et al. 2006), large 12CO(1-0) sizes (Ivison et al. 2010a, 2011; Thomson et al. 2012; Hodge et al. 2012), lack of silicate absorption in mid- infrared spectra (Men´endez-Delmestre et al. 2009), and less reddened broad-band mid-infrared colours (Hainline et al. 2009). High [CII]/far-infrared ratios (Stacey et al. 2010; Ivi- son et al. 2010b; S12) add to this mounting evidence that star-formation in SMGs takes place in a region larger than the compact nuclear starbursts of local ULIRGs. |
Processed_Physical_Conditions_of_the_Gas_in_an_ALMA_[CII]-id.txt | mological parameters of ΩM = 0.27, ΩΛ = 0.73, and a Hub- ble constant of 71 km s−1 Mpc−1 throughout this paper. |
Processed_Physical_Conditions_of_the_Gas_in_an_ALMA_[CII]-id.txt | The 12CO(2–1) line (νrest = 230.538 GHz) in ALESS65.1 (RA(J2000) = 03 32 52.26, Dec(J2000) = −27 35 26.3) (S12) was observed over a period of four consecutive nights, 9 – 12 August 2012, with the Australia Telescope Compact Ar- ray (ATCA), using the Compact Array Broadband Backend (CABB). The array was in the most compact five-antenna configuration, H75, which has a maximum baseline of 89m and two antennas set along a northern spur. The hybrid configuration allows good (u, v) coverage to be obtained for integrations less than the full 12 hour synthesis. We centered the 7mm IF1 receiver on 42.343 GHz, the expected frequency of the 12CO(2–1) line emission given the [CII] redshift of z = 4.4445 derived by S12. The 2GHz wide bandwidth of CABB results in a frequency coverage of 41.3 to 43.3 GHz, covering 12CO(2–1) emission between z = 4.32 – 4.58. The weather was good to average, with rms atmospheric path length variations of 100 to 400 microns throughout the run, as measured on the 230m baseline ATCA Seeing Monitor (Middelberg et al. 2006). The system temperature was 140 to 250 K throughout the four nights. Weather and atmo- spheric conditions can induce temporal fluctuations across the wide CABB band, so, following Emonts et al. (2011), a bandpass calibration scan was acquired at the beginning, middle and end of each 8 hour night. Phase and amplitude calibration information was acquired with 2 minute scans on PKS 0346−279 every 10 minutes and pointing checks performed on the same source every hour. Flux density cal- ibration was performed on Uranus at the beginning of the nights, when it was at an elevation of roughly 55 degrees. The uncertainty in the flux density calibration using the standard MIRIAD model of Uranus is estimated to be 30% (Emonts et al. 2011). |
Processed_Physical_Conditions_of_the_Gas_in_an_ALMA_[CII]-id.txt | The data were calibrated, mapped and analysed using the standard MIRIAD (Sault & Killeen 1999) and KARMA (Gooch 1996) packages. The synthesized beam from natural weighting is 14.0 × 9.9 arcsec. A total of about 20 hours on-source integration time was obtained over the 4 × 8 hour nights. ALESS65.1 was not detected in the 42.3 GHz con- tinuum map from the full CABB band, which achieves an rms noise level of 11 µJy beam−1. |
Processed_Physical_Conditions_of_the_Gas_in_an_ALMA_[CII]-id.txt | Figure 1. 12CO(2-1) spectrum of ALESS65.1 binned into 100 km s−1 channels and extracted at the ALMA position. The red dashed line shows the ALMA [CII] spectrum from S12, binned to a similar channel width (120 km s−1) and with the flux density divided by 15 for clarity. The blue dotted line shows the spectrum binned into 200 km s−1 for maximum sensitivity. We conclude that no line was detected to a 3σ limit of 0.24 mJy beam−1 per 200 km s−1 channel. |
Processed_Physical_Conditions_of_the_Gas_in_an_ALMA_[CII]-id.txt | FWHM identical to the [CII] emission line (470 km s−1, S12) and adopting a 3σ limit, the 12CO(2–1) line intensity of ALESS65.1 is ICO(2−1) < 0.11 Jy km s−1. |
Processed_Physical_Conditions_of_the_Gas_in_an_ALMA_[CII]-id.txt | The observed and derived properties of ALESS65.1 are sum- marised in Table 1. To estimate the cold molecular gas con- tent in ALESS65.1 we calculate upper limits to the line lu- minosity and total cold gas (H2 + He) mass from the CO(2– 1) flux density limit. Following Solomon and Vanden Bout (2005), the line luminosity upper limit is L(cid:48) CO(2−1) < 2.2 × 1010 K km s−1 pc2. If we assume the gas is thermalised (i.e. intrinsic brightness temperature and line luminosities are independent of J transition), so L(cid:48) CO(1−0), and a CO-to-H2 conversion factor α = 0.8 M(cid:12) (K km s−1 pc2)−1, which is appropriate for ULIRGs (e.g. Downes & Solomon 1998, but see Bothwell et al. 2012), the upper limit on the total cold gas mass is Mgas < 1.7 × 1010 M(cid:12). This is consistent with the gas mass found in other z > 4 SMGs (Schinnerer et al. 2008; Daddi et al. 2009a; Coppin et al. 2010; Walter et al. 2012) and the typical [CII]/Mgas ratio at high redshift (S12). |
Processed_Physical_Conditions_of_the_Gas_in_an_ALMA_[CII]-id.txt | The total baryonic mass of ALESS65.1 can be derived by combining the gas and stellar mass estimates for the sys- tem. The stellar mass of the system was estimated from the rest-frame absolute H-band magnitude to be M∗ ∼ 9 × 1010 M(cid:12) (S12), so the gas fraction is modest with Mgas/M∗ (cid:46) 0.2. The total baryonic mass Mbary = Mgas+M∗ is ∼ 9 – 11×1010 M(cid:12). This is consistent with the dynamical mass for the sys- tem, based on the [CII] linewidths and spatial extent, of Mdyn sin2(i) ∼ (3.4 ± 1.8) × 1010 M(cid:12) (S12). |
Processed_Physical_Conditions_of_the_Gas_in_an_ALMA_[CII]-id.txt | Next we examine the physical conditions of the gas in ALESS65.1 using the [CII] detection and 12CO(2–1) limit. The L[CII]/LFIR versus LCO(1−0)/LFIR diagram is a pow- erful diagnostic as these two ratios are sensitive to gas density n and the incident far-ultraviolet (FUV) flux G0 (Stacey et al. 1991). Figure 2 shows ALESS65.1 compared with other low and high redshift galaxies, and solar metal- lically PDR model curves (Kaufman et al. 1999). This dia- gram can be used to roughly estimate both n and G0 for a galaxy, but there are some caveats, as outlined in De Breuck et al. (2011). These are: (i) the 12CO luminosity is that of the ground state rotational line, (ii) [CII] emis- sion from the diffuse ionised medium and cosmic-ray-heated gas is assumed to be small compared to that from PDRs, and (iii) AGN and their related X-Ray Dissociation Regions (XDRs) are assumed to not contribute significantly to the FIR and [CII] luminosity. To be consistent with both De Breuck et al. (2011) and Stacey et al. (2010) in Figure 2 we assume LCO(2−1)/LCO(1−0) = 7.2, which is 90% of its value if the gas was fully thermalised and optically thick. Cos- mic ray rates are greater in starbursts compared to normal galaxies but this does not seem to result in higher [CII]/CO ratios, so cosmic ray ionization does not appear to domi- nate the [CII]/CO ratio (De Breuck et al. 2011), at least for local galaxies. ALESS65.1 is not detected in the 250ks Chandra X-Ray observations of this region (Lehmer et al. 2005), so it is not a luminous QSO (L3−44keV (cid:46) 2–3×1044 erg s−1, for NH = 0 – 1023.5 cm−2). The maximum likely AGN contribution to the FIR luminosity is estimated from a multicomponent (AGN and starburst) fit to the FIR us- ing Decompir software (Mullaney et al. 2011) and following the method described by (Seymour et al. 2012). We use the 24 µm flux density upper limit to constrain the short wave- length part of the IR SED and find the AGN contribution to the total FIR luminosity is (cid:46)10%. ALESS65.1 therefore appears to be dominated by star-formation processes and the AGN contribution to [CII] and LFIR is minimal. |
Processed_The_Strong_Isospin-Breaking_Correction_for_the_Glu.txt | Kim Maltman† Department of Mathematics and Statistics, York University, 4700 Keele St., Toronto, Ontario, Canada M3J 1P3 and Special Research Centre for the Subatomic Structure of Matter, University of Adelaide, Australia 5005. |
Processed_The_Strong_Isospin-Breaking_Correction_for_the_Glu.txt | The strong isospin-breaking correction, Ωst, which appears in estimates of the Standard Model value for the direct CP-violating ratio ǫ′/ǫ, is evaluated to next-to-leading order (NLO) in the chiral expansion using Chiral Perturbation Theory. The relevant linear combinations of the unknown NLO CP-odd weak low-energy constants (LEC’s) which, in combination with 1-loop and strong LEC con- tributions, are required for a complete determination at this order, are estimated using two different It is found that, to NLO, Ωst = 0.08 ± 0.05, significantly reduced from the “standard” models. value, 0.25 ± 0.08, employed in recent analyses. The potentially significant numerical impact of this decrease on Standard Model predictions for ǫ′/ǫ, associated with the decreased cancellation between gluonic penguin and electroweak penguin contributions, is also discussed. |
Processed_The_Strong_Isospin-Breaking_Correction_for_the_Glu.txt | Chiral perturbation theory (ChPT) provides a natural framework for the calculation of Ωst since it ensures that all contributions of a given chiral order may be obtained in a computationally straightforward manner. The complete set of NLO contributions is a sum of NLO strong LEC, one-loop, and NLO weak LEC contributions, each of which is separately renormalisation-scale-dependent, divergent, and therefore unphysical. The sum of these contributions is, however, necessarily finite and scale-independent. The cancellation of both divergences and scale-dependence in the final result provides a highly non-trivial check of the explicit calculations. The set of Feynmann graphs to be evaluated is shown in Fig. 1, where Fig. 1(a) represents the LO, Fig. 1(c) the NLO strong LEC, Figs. 1(b) and 1 (d)-(g) the one-loop, and Fig. 1(h) the NLO weak LEC contributions, respectively. In Ref. [16] only the NLO contributions associated with external line dressing (Figs. 1(b,c)) were considered, while Ref. [18] examined only the contribution of Fig. 1(h). |
Processed_The_Strong_Isospin-Breaking_Correction_for_the_Glu.txt | In Refs. [19,20] it was shown that, in the CP-even sector, the inclusion of one-loop isospin-breaking corrections in ππ decay amplitudes acts to decrease the magnitude of the octet leakage contribution to the ∆I = 3/2 amplitude. K This decrease ensures that the ∆I = 1/2 rule enhancement obtained in an IC analysis is accurate to better than 10%. (p2δm)) determination of Ωst in ChPT, and show that the same In the present paper we obtain a complete NLO ( qualitative situation holds in the CP-odd sector. This means, according to Eq. 4, that one should expect a further reduction of Ωst beyond that associated with the NLO strong mixing effects studied in Ref. [16]. The absence of reliable values for the CP-odd weak LEC’s is the crucial stumbling block on the way to an accurate numerical result for Ωst at NLO, the NLO strong LEC and loop corrections being, as we will see below, well-determined. In the present work we appeal to two models (the weak deformation model, and the chiral quark model) of the contribution of the gluonic penguin operator, Q6, to the CP-odd weak LEC’s in order to probe the probable scale of the model-dependence in our estimates of Ωst. It is important to emphasize that, while the estimates of the weak LEC contributions to Ωst are model-dependent, the one-loop and strong LEC contributions (discussed below) are, though scale-dependent, model-independent. As we will see, the combination of these model-independent NLO contributions is, at typical hadronic scales, rather large, and negative, suggesting a significant reduction of Ωst as compared to the conventional value. |
Processed_The_Strong_Isospin-Breaking_Correction_for_the_Glu.txt | The rest of the paper is organised as follows: In the next section we briefly review the chiral Lagrangian approach to the calculation of non-leptonic kaon decay amplitudes and discuss the models employed for the relevant NLO weak LEC combinations. In Section III we present our numerical results for Ωst. The impact of our findings on theoretical estimates for the value of ǫ′/ǫ in the Standard Model are discussed and conclusions are presented in Section IV. |
Processed_The_Strong_Isospin-Breaking_Correction_for_the_Glu.txt | where U = exp(iλ chiral limit. The superscripts represents the LO CP-odd octet weak coupling strength. |
Processed_The_Strong_Isospin-Breaking_Correction_for_the_Glu.txt | estimates are available, the chiral quark model (χQM) as implemented by the Trieste group in Refs. [8–10] for which all the necessary ingredients required to calculate the NLO weak LEC contributions to ImA0 and ImδA2 are readily available, and the “scalar saturation model” of Ref. [18], which combines the factorisation approximation with the assumption of scalar meson exchange saturation of the relevant strong LEC’s. |
Processed_The_Strong_Isospin-Breaking_Correction_for_the_Glu.txt | In another approach, the hadronic matrix elements of the four-quark operators relevant for non-leptonic kaon decay have been estimated in the χQM by the Trieste group [8–10]. In particular, expressions for the gluonic penguin contribution to the LO weak LEC’s were obtained in the second of Refs. [10], and preliminary estimates of the contribution to NLO weak LEC’s were given in Ref. [8]. As pointed out by the authors of Ref. [8], however, the latter expressions contain errors and are not to be used [28]. We have used the updated results of Ref. [9], which give the contributions of Q6 to A0 to NLO, to extract the CP-odd weak LEC combination which enters ImA0. To obtain the IB = (√3/2)δA2 combination of NLO weak LEC’s entering δA2, we have calculated the matrix element in the χQM using the formalism described in Ref. [9], together with the corrected results for the basic ingredients required to compute such matrix elements in the model given there. The model parameters which enter these estimates are the constituent quark mass, M , the vacuum quark condensate, and the gluon condensate. These parameters were constrained to reproduce the ∆I = 1/2 Rule in the CP-even sector in Ref. [9] using a matching scale Λχ = 0.8 GeV (the scale at which the scale-dependence of the chiral loops and that of the short-distance expressions is roughly −10 GeV)3, and matched in the model). The fitted parameters have the values M = 0.2 0.004 GeV)4. The resulting values for the NLO weak LEC combinations are presented in the αsGG/π h first column of Table I. Numerical estimates for the NLO weak LEC contributions to ImA0 and ImδA2 are presented in the next section. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | Self-Supervised Learning (SSL) has been shown to learn useful and information- preserving representations. Neural Networks (NNs) are widely applied, yet their weight space is still not fully understood. Therefore, we propose to use SSL to learn hyper-representations of the weights of populations of NNs. To that end, we introduce domain specific data augmentations and an adapted attention architecture. Our empirical evaluation demonstrates that self-supervised representation learning in this domain is able to recover diverse NN model characteristics. Further, we show that the proposed learned representations outperform prior work for predicting hyper-parameters, test accuracy, and generalization gap as well as transfer to out-of-distribution settings. Code and datasets are publicly available1. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | This work investigates populations of Neural Network (NN) models and aims to learn representations of them using Self-Supervised Learning. Within NN populations, not all model training is successful, i.e., some overfit and others generalize. This may be due to the non-convexity of the loss surface during optimization [Goodfellow et al., 2015], the high dimensionality of the optimization space, or the sensitivity to hyperparameters [Hanin and Rolnick, 2018], which causes models to converge to different regions in weight space. What is still not yet fully understood, is how different regions in weight space are related to model characteristics. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | Previous work has made progress investigating characteristics of NN models, e.g by visualizing learned features [Zeiler and Fergus, 2014, Karpathy et al., 2015]. Another line of work compares the activations of pairs of NN models [Raghu et al., 2017, Morcos et al., 2018, Kornblith et al., 2019]. Both approaches rely on the expressiveness of the data, and are, in the latter case, limited to two models at a time. Other approaches predict model properties, such as accuracy, generalization gap, or hyperparameters from the margin distribution [Yak et al., 2019, Jiang et al., 2019], graph topology features [Corneanu et al., 2020] or eigenvalue decomposition of the weight matrices [Martin and Mahoney, 2019]. In a similar direction, other publications propose to investigate populations of models and to predict properties directly from their weights or weight statistics in a supervised way [Unterthiner et al., 2020, Eilertsen et al., 2020]. However, these manually designed features may not fully capture the latent model characteristics embedded in the weight space. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | Figure 1: An overview of the proposed self-supervised representation learning approach. I. Popula- tions of trained NNs form model zoos; each model is transformed in a vectorized form of its weights. II. Hyper-representations are learned from the model zoos using different augmentations, architec- tures, and Self-Supervised Learning tasks. III. Hyper-representations are evaluated on downstream tasks which predict model characteristics. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | Therefore, our goal is to learn task-agnostic representations from populations of NN models able to reveal such characteristics. Self-Supervised Learning (SSL) is able to reveal latent structure in complex data without the need of labels, e.g., by compressing and reconstructing data [Goodfellow et al., 2016, Kingma and Welling, 2014]. Recently, a specific approach to SSL called contrastive learning has gained popularity [Misra and van der Maaten, 2019, Chen et al., 2020a,b, Grill et al., 2020]. Contrastive learning leverages inherent symmetries and equivariances in the data, allows to encode inductive biases and thus structure the learned representations. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | In this paper, we propose a novel approach to apply SSL to learn representations of the weights of NN populations. We learn representations using reconstruction, contrast, and a combination of both. To that end, we adapt a transformer architecture to NN weights. Further, we propose three novel data augmentations for the domain of NN weights. We introduce structure preserving permutations of NN weights as augmentations, which make use of the structural symmetries within the NN weights that we find necessary for learning generalizing representations. We also adapt erasing [Zhong et al., 2020] and noise [Goodfellow et al., 2016] as augmentations for NN weights. We evaluate the learned representations by linear-probing for the generating factors and characteristics of the models. An overview for our learning approach is given in Figure 1. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | To validate our approach, we perform extensive numerical experiments over different populations of trained NN models. We find that hyper-representations2 can be learned and reveal the characteristics of the model zoo. We show that hyper-representations have high utility on tasks such as the prediction of hyper-parameters, test accuracy, and generalization gap. This empirically confirms our hypothesis on meaningful structures formed by NN populations. Furthermore, we demonstrate improved performance compared to the state-of-the-art for model characteristic prediction and outlay the advantages in out-of-distribution predictions. For our experiments, we use publicly available NN model zoos and introduce new model zoos. In contrast to [Unterthiner et al., 2020, Eilertsen et al., 2020], our zoos contain models with different initialization points and diverse configurations, and include densely sampled model versions during training. Our ablation study confirms that the various factors for generating a population of trained NNs play a vital role in how and which properties are recoverable for trained NNs. In addition, the relation between generating factors and model zoo diversity reveals that seed variation for the trained NNs in the zoos is beneficial and adds another perspective when recovering NNs’ properties. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | Model Zoo We denote as D a data set that contains data samples with their corresponding labels. We denote as λ the set of hyper-parameters used for training, (e.g., loss function, optimizer, learning rate, weight initialization, batch-size, epochs). We define as A the specific NN architecture. Training under different prescribed configurations {D, λ, A} results in a population of NNs which we refer to as model zoo. We convert the weights and biases of all NNs of each model into a vectorized form. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | in analogy to HyperNetworks [Ha et al., 2016], which are trained to generate weights for larger NN models. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | In the resulting model zoo W = {w1, ...., wM }, wi denotes the flattened vector of dimension N , representing the weights and biases for one trained NN model. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | Augmentations. Data augmentation generally helps to learn robust models, both by increasing the number of training samples and preventing overfitting of strong features [Shorten and Khoshgoftaar, 2019]. In contrastive learning, augmentations can be used to exploit domain-specific inductive biases, e.g., know symmetries or equivariances [Chen and Li, 2020]. To the best of our knowledge, augmentations for NN weights do not yet exist. In order to enable self-supervised representation learning, we propose three methods to augment individual instances of our model zoos. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | where ˆWl+1 = Wl+1(Pl)T, ˆWl = PlWl and ˆbl = Plbl are the permuted weight matrices and bias vector, respectively. The equivalences hold not only for the forward pass, but also for the backward pass and weight update.3 The permutation can be extended to kernels of convolution layers. The permutation augmentation differs significantly from existing augmentation techniques. As an analogy, flips along the axis of images are similar, but specific instances from the set of possible permutations in the image domain. Each permutable layer with dimension Nl, has Nl! different permutation matrices, and in total there are (cid:81) l Nl! distinct, but equivalent versions of the same NN. While new data can be created by training new models, the generation is computationally expensive. The permutation augmentation, however, allows to compute valid NN samples at almost no computational cost. Empirically, we found the permutation augmentation crucial for our learning approach. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | In computer vision and natural language processing, masking parts of the input has proven to be helpful for generalization [Devlin et al., 2019]. We adapt the approach of random erasing of sections in the vectorized forms of trained NN weights. As in [Zhong et al., 2020], we apply the erasing augmentation with a probability p to an area that is randomly chosen with a lower and upper bounds blow and bup. In our experiments, we set p = 0.5, blow = 0.03, bup = 0.3 and erase with zeros. Adding noise augmentation is another way of altering the exact values of NN weights without overly affecting their mapping, and has long been used in other domains [Goodfellow et al., 2016]. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | With this work, we propose to learn representations of structures in weight space formed by popu- lations of NNs. We evaluate the representations by predicting model characteristics. A supervised learning approach has been demonstrated [Unterthiner et al., 2020, Eilertsen et al., 2020]. [Unterthiner et al., 2020] find that statistics of the weights (mean, var and quintiles) are superior to the weights to predict test accuracy, which we empirically confirm in our results. However, we intend to learn representations of the weights, that contain rich information beyond statistics. ‘Labels’ for NN models can be obtained relatively simply, yet they can only describe predefined characteristics of a model instance (e.g., accuracy) and so supervised learning may overfit few features, as [Unterthiner et al., 2020] show. Self-supervised approaches, on the other hand, are designed to learn task-agnostic representations, that contain rich and diverse information and are exploitable for multiple downstream tasks [LeCun and Misra, 2021]. Below, we present the used architectures and losses for the proposed self-supervised representation learning. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | Figure 2: Trained NN weights are the input sequence to a transformer. Left. Each element in the sequence represents the weight that connects two neurons at two different layers. Right. Each element in the sequence represents the set of weights related to one neuron. |
Processed_Hyper-Representations:_Self-Supervised_Representat.txt | Figure 3: Multi-head attention-based encoder. Left. Regular sequence-to-sequence translation, each element of the output sequence is used. Right. An additional compression token is added to the sequence. From the output sequence, only the compression token is taken. |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | high-z DLAs is their typically-low metallicity [Z/H], imply- ing fewer routes for gas cooling. If high Ts values in DLAs arise due to low absorber metallicities, one would expect an anti-correlation between Ts and [Z/H], with low-Ts DLAs hav- ing high metallicities, and vice-versa (Kanekar & Chengalur 2001). We report here the detection of the predicted anti- correlation between Ts and [Z/H], supporting the conclusion that HI in high-z DLAs is predominantly in the warm neutral medium. |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | Notes: The Ts values have been re-computed uniformly by Kanekar et al. (in prep.). References for [Z/H] values are mostly to literature compilations, which contain the original references; all values are scaled to the solar abundances of Lodders (2003). (1) Kulkarni et al. (2005); (2) Kanekar et al. (in prep.); (3) Boisse et al. (1998); (4) Junkkarinen et al. (2004); (5) Wolfe et al. (2008); (6) Nestor et al. (2008); (7) Prochaska et al. (2007); (8) Ledoux et al. (2006); (9) Ellison et al. (2008); (10) Akerman et al. (2005); (11) Prochaska et al. (2003b); (12) Ellison et al. (2001). |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | Note that the sample does not include four systems where the radio emission is clearly extended on scales >> 1′′, and the covering factor is likely to be low, f << 1; these are the HI 21cm absorbers at z ∼ 0.437 towards 3C196 (Briggs et al. 2001) and z ∼ 0.656 towards 3C336 (Curran et al. 2007a) (where the HI 21cm absorption arises towards extended lobes, with little radio flux density associated with the quasar core), and the HI 21cm non-detections at z ∼ 1.3911 to- wards QSO 0957+561 and z ∼ 1.4205 towards PKS 1354+258 (Kanekar & Chengalur 2003). We have also excluded “asso- ciated” DLAs, lying within ∼ 3000 km s- 1 of the quasar, as conditions in these absorbers could be affected by their prox- imity to an active galactic nucleus. |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | could not rule out the effects of unknown covering factors. The results of the present Letter are based on new HI 21cm data on 11 absorbers, and VLBI estimates of the DLA cover- ing factor for most systems, and yield the first clear evidence for a relation between the metallicity and the HI temperature distribution in the absorbers. |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | with measurements of both Ts and [Z/H]. The dashed line shows the linear fit to the relation between [Z/H] and Ts. |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | at ∼ 3.1σ significance, showing that it does not arise from incorrect metallicity estimates in these absorbers. |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | A linear regression analysis was used to obtain the best- fit relation between Ts and [Z/H], using the ten systems with measurements of both Ts and [Z/H] (i.e. excluding all limits), all of which also have covering factor estimates; these are shown in Fig. 1[B]. The BCES(Y/X) estimator (Akritas & Bershady 1996) was used for this purpose, treat- ing [Z/H] as the independent variable X; this method takes into account measurement errors on both variables, as well as the possibility that these errors are correlated (which ap- plies here, as Ts and [Z/H] are both derived from NHI). We obtain Ts = (- 0.68 ± 0.17) × [Z/H] + (2.13 ± 0.21); the fit is shown as a dashed line in Figs. 1[A] and [B]. Exclud- ing the DLAs towards 0235+164 and 0827+243 yields Ts = (- 0.77 ± 0.40) × [Z/H]+ (2.02 ± 0.48), from eight DLAs with measurements of Ts and [Z/H]. |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | We also tested for a correlation between the dust depletion factor [Z/Fe] and Ts, using the 21 systems with estimates of both quantities. The anti-correlation between [Z/Fe] and Ts has ∼ 2.2σ significance, reducing to ∼ 1.6σ significance on excluding the possibly-unreliable X-ray metallicity estimate in the DLA towards 0235+164. We thus do not find significant evidence for a relation between [Z/Fe] and Ts, although this sub-sample contains few DLAs with low spin temperatures. |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | Curran et al. (2007b) note that an observed anti-correlation between Ts and metallicity might arise due to unknown low covering factors. We hence estimated the BHK statistic for the sub-sample of 20 DLAs with covering factor estimates (Kanekar et al. 2009a). The anti-correlation between Ts and [Z/H] is then detected at ∼ 3σ significance, with a probability of ∼ 0.003 of chance occurrence. Low covering factors are thus unlikely to be the cause of the anti-correlation between Ts and [Z/H]. |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | locity spread and metallicity; Ledoux et al. 2006). For ex- ample, Curran et al. (2007b) argue that an observed correla- tion between Ts and [Z/H] might arise due to a relation be- tween the velocity spread of HI 21cm absorption ∆V21 and the rest equivalent width of the MgIIλ2796 line, W λ2796 , each tracing the dynamics of the absorber. They note that Ts ∝ NHI/ (cid:2)R τ21dV (cid:3) ≈ NHI/ (cid:2)τ21,max∆V21(cid:3), suggesting that Ts ∝ 1/∆V21. W λ2796 is known to correlate with metallicity (Murphy et al. 2007), so, if ∆V21 also correlates with W λ2796 , the three relations could yield an observed anti-correlation be- tween Ts and [Z/H]. However, the HI column density in DLAs does not correlate with either [Z/H] or W λ2796 (e.g. Rao et al. 2006), and the DLAs of Table 1 have NHI values extending over ≈ 1.5 dex. If the Ts–[Z/H] relation arises due to an un- derlying relation between ∆V21 and W λ2796 , this spread in NHI values would imply that the Ts–[Z/H] relation would be weaker than the relation between ∆V21 and W λ2796 (and vice- versa, if the Ts–[Z/H] relation is the “primary” relation). In other words, the “primary” relation would have the highest statistical significance in samples of similar size, as uncor- related (and variable) parameters like NHI would dilute the significance of any derived relations. |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | versely, the mass-metallicity relation implies that high-mass DLAs would have high metallicities, along with high ther- mal pressures. Such galaxies would have significant CNM fractions, with most sightlines yielding low spin tempera- tures. These effects would yield the anti-correlation be- tween Ts and [Z/H] detected here, due to the paucity of cooling routes in low-metallicity galaxies (Norman & Spaans 1997; Kanekar & Chengalur 2001; Kanekar & Briggs 2004). The anti-correlation between Ts and [Z/H] is thus consis- tent with the presence of a mass-metallicity relation in DLAs (Ledoux et al. 2006; Prochaska et al. 2008); low-metallicity DLAs are likely to have high Ts values due to their low CNM fractions. |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | While the relation between Ts and [Z/H] is consistent with the presence of a mass-metallicity relation in DLAs, it is also possible that the Ts–[Z/H] relation is a local one, arising due to line-of-sight issues. For disk galaxies, the cross-section for DLA incidence is largest for sightlines through the out- skirts of the galaxy. If high-z DLAs have steep metallic- ity gradients, such sightlines would typically encounter low metallicities (e.g. Zwaan et al. 2005). The lack of local cooling routes could then result in low CNM fractions, and high Ts values along these sightlines. Conversely, sightlines through the central regions of large disk galaxies would typ- ically have high metallicities and high CNM fractions. Such line-of-sight effects could thus also yield the anti-correlation of Fig. 1. While such steep metallicity gradients have not been seen in local galaxies and low-z DLAs (typical gradients are 0.04 dex kpc- 1; Chen et al. 2005; Bresolin et al. ∼ - 0.03 - 2009), they are not ruled out in high-z DLAs. We note, in passing, that the z ∼ 0.524 DLA towards 0827+243 has one of the largest DLA impact parameters (∼ 27 kpc; Steidel et al. 2002), and yet has both a low Ts (330 K) and a high metallic- ity, [Z/H]= - 0.6. |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | Fig. 1[A] shows that low Ts values are only found in DLAs with [Z/H]& - 0.6. Only 13 of the 153 DLAs at z > 1.6 in the sample of Prochaska et al. (2007) have such high metal- licities; the majority of high-z DLAs have [Z/H]< - 1.0. The anti-correlation between Ts and [Z/H] then implies that most high z DLAs have high Ts values. Conversely, six of seven- teen DLAs at z . 1 have [Z/H]≥ - 0.6 (e.g. Prochaska et al. 2003a; Kulkarni et al. 2005). While the number of metallicity measurements at z < 1 is yet small, the present data indicate a higher fraction of high-metallicity DLAs at low redshifts. If so, the Ts–[Z/H] anti-correlation implies that the fraction of DLAs with high CNM fraction should also be larger at z . 1. This is consistent with the observed increase in the detection rate of HI 21cm absorption in DLAs at z . 1 (Kanekar et al. 2009b). |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | and are instead likely to reflect the underlying gas temper- ature distribution. The majority of DLAs at z > 1.6 have [Z/H]< - 1, implying that most of the HI in DLAs must be in the warm phase, with small CNM fractions. The observed anti-correlation between Ts and [Z/H] is consistent with inde- pendent evidence for the presence of a mass-metallicity rela- tion in DLAs. The majority of high-z DLAs are likely to be galaxies of low mass and metallicity, with most of the neutral gas in the warm phase. |
Processed_A_metallicity-spin_temperature_relation_in_damped_.txt | We thank Sara Ellison and an anonymous referee for com- ments on an earlier draft. Part of this work was carried out during a visit by NK to ESO, Santiago, under an ESO Vis- iting Fellowship; he thanks ESO for support and hospitality, and is also grateful for support from the Max-Planck Society and the Alexander von Humboldt Foundation. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.