SHAT: A Novel Asynchronous Training Algorithm That Provides Fast Model Convergence in Distributed Deep Learning

The recent unprecedented success of deep learning (DL) in various fields is underlied by its use of large-scale data and models. Training a large-scale deep neural network (DNN) model with large-scale data, however, is time-consuming. To speed up the training of massive DNN models, data-parallel distributed training based on the parameter server (PS) has been widely applied. In general, a synchronous PS-based training suffers from the synchronization overhead, especially in heterogeneous environments. To reduce the synchronization overhead, asynchronous PS-based training employs the asynchronous communication between PS and workers so that PS processes the request of each worker independently without waiting. Despite the performance improvement of asynchronous training, however, it inevitably incurs the difference among the local models of workers, where such a difference among workers may cause slower model convergence. Fro addressing this problem, in this work, we propose a novel asynchronous PS-based training algorithm, SHAT that considers (1) the scale of distributed training and (2) the heterogeneity among workers for successfully reducing the difference among the local models of workers. The extensive empirical evaluation demonstrates that (1) the model trained by SHAT converges to the higher accuracy up to 5.22% than state-of-the-art algorithms, and (2) the model convergence of SHAT is robust under various heterogeneous environments.

CLIMATE REGULATING ECOSYTEM SERVICES OF THE FOREST LANDSCAPE IN THE UKRAINIAN CARPATHIANS

Ecosystems provide multiple services for humans. Among them, a group of supporting and regulating ecosystem services is often less recognized by people as benefit and has been less studied by researchers. Amid various manifestations of climate change, more attention has been paid to particular subset of this group of services called climate regulating. Despite these there still few quantitative studies in this field. Trying to fill this research gap we conducted a study aimed at exploration of relation between climate regulating ecosystem services and their spatial determinants in the forest landscape within Ukrainian Carpathians. For that we chose the territory within Rakhiv and Tsiachiv districts in Transcarpathian region which represents all diversity of forest mountain ecosystems. For this study we used information on land surface temperature (LST) extracted from Landsat 8 thermal band for summer season of 2015. In order to account for vertical thermal gradient in mountains the LST data underwent normalization and in further analysis a dependent variable we employed normalized LST (nLST). Set of independent variables included geomorphometric indicators (altitude, slope, aspect, TPI) and data on forest cover (disturbance, density, dominant species, and disturbance in the neighborhood). For key study area of Velykyi watershed of 4059 ha we additionally used data on forest biomass and tree age. In general, all forest ecosystems in present research have been divided into three distinct classes – “natural”, “disturbed” and “other” forests. Using boosted regression trees method we built three statistical models for each of the forest classes called “global” models. Also we developed 12 “local” models that showed the link between nLST and analyzed independent variables within each altitudinal bioclimatic zone with considering also forest class. Three separate statistical models have been built for each of the forest classes for key study area. Our results suggest that both maximum and mean values of nLST within particular altitudinal bioclimatic zone are the lowest in “natural” forests and the highest in “disturbed” ones.. The statistical model performance based on the variance explained indicator ranged from 32 to 74 %, whilst for models for key study area it was between 77 and 89 %. The set of influential variables for different forest classes varied substantially, but the most often they included aspect, forest density and elevation despite of normalization applied before. In models created for class “disturbed” forests between 19 and 35 % of all explained variance has been contributed by variable indicating time of disturbance. In “local” models for class “natural” forests we revealed gradual decrease of influence of the geomorphometric indicators (elevation, slope, and TPI) when move from warmer to cooler altitudinal zones while for topographic aspect and forest density the trends were just the opposite. In case of key study area a wood stock and tree age variables along with elevation and aspect were amongst the most influential ones. We can conclude that depending on the state of naturalness of forest ecosystems they have different climate regulating potential which might be severely depleted by human and natural disturbances. Keywords: forest landscape, ecosystem services, remote sensing, climate regulation, climate change, Landsat satellite images, Ukrainian Carpathians.

Less entanglement exhibiting more nonlocality with noisy measurements

The Clauser–Horne–Shimony–Holt (CHSH) inequality test is widely used as a mean of invalidating the local deterministic theories. Most attempts to experimentally test nonlocality have presumed unphysical idealizations that do not hold in real experiments, namely, noiseless measurements. We demonstrate an experimental violation of the CHSH inequality that is free of idealization and rules out local models with high confidence. We show that the CHSH inequality can always be violated for any nonzero noise parameter of the measurement. Intriguingly, less entanglement exhibits more nonlocality in the CHSH test with noisy measurements. Furthermore, we theoretically propose and experimentally demonstrate how the CHSH test with noisy measurements can be used to detect weak entanglement on two-qubit states. Our results offer a deeper insight into the relation between entanglement and nonlocality.

Stepped-ness in medium density housing: The stepped dwellings

<p>Stepped-ness in Medium Density Housing investigates a new form of design, where the site is organised according to different conditions of public, common, shared and private spaces. Stepped-ness is used as a technique for controlling relations at a range of scales and intimacies —from urban to interior— and as a tool for creating continuity of public to private, inside to outside and building to landscape. As a result, circulation and dwelling become integrated as part of a stepped morphology in which higher density living is able to accommodate both a desire for privacy, and a connection with neighbours. The typical detached New Zealand house reinforces the nuclear family as unit and precludes the extended family. Local models of medium density housing replicate these conventions and continue to deliver autonomous and identifiably singular buildings defined by lot and footprint size. Challenging these conventions, the architectural hardware of this proposal allows boundaries to be redefined according to the preferred size and configuration of a variety of household types. The identity of ‘home’ is less determined by size, and more by relations – within the household and between dwelling and public realm. Insistently organisational, mat-building explores relational distances. The stepped-mat is examined in terms of the design of mediatory devices and ancillary spaces as the primary element of space planning. This was inspired by Atelier Bow Wow, who is known for creating, “not an architecture of spaces, but an architecture of relationships,” (Fujimori, 2010, p. 128). Initially concerned with the step as a tool for design, this thesis developed to consider the stepped-mat as a way of both controlling and designing space at a range of scales. How then, could the kinds of interstitial and circulation spaces between dwellings be exploited for incidental meetings between neighbours, to sit outside and read a book, or bring some daylight into a corridor (Barker & Simons, 2012, p.155). Designed to both transport and accommodate us, the stairs’ behaviour in the breakdown of vertical and horizontal space drives this design, and forms new relations between households.</p>

Stepped-ness in medium density housing: The stepped dwellings

<p>Stepped-ness in Medium Density Housing investigates a new form of design, where the site is organised according to different conditions of public, common, shared and private spaces. Stepped-ness is used as a technique for controlling relations at a range of scales and intimacies —from urban to interior— and as a tool for creating continuity of public to private, inside to outside and building to landscape. As a result, circulation and dwelling become integrated as part of a stepped morphology in which higher density living is able to accommodate both a desire for privacy, and a connection with neighbours. The typical detached New Zealand house reinforces the nuclear family as unit and precludes the extended family. Local models of medium density housing replicate these conventions and continue to deliver autonomous and identifiably singular buildings defined by lot and footprint size. Challenging these conventions, the architectural hardware of this proposal allows boundaries to be redefined according to the preferred size and configuration of a variety of household types. The identity of ‘home’ is less determined by size, and more by relations – within the household and between dwelling and public realm. Insistently organisational, mat-building explores relational distances. The stepped-mat is examined in terms of the design of mediatory devices and ancillary spaces as the primary element of space planning. This was inspired by Atelier Bow Wow, who is known for creating, “not an architecture of spaces, but an architecture of relationships,” (Fujimori, 2010, p. 128). Initially concerned with the step as a tool for design, this thesis developed to consider the stepped-mat as a way of both controlling and designing space at a range of scales. How then, could the kinds of interstitial and circulation spaces between dwellings be exploited for incidental meetings between neighbours, to sit outside and read a book, or bring some daylight into a corridor (Barker & Simons, 2012, p.155). Designed to both transport and accommodate us, the stairs’ behaviour in the breakdown of vertical and horizontal space drives this design, and forms new relations between households.</p>

E-scooter’s Availability and Social Equity in Minneapolis, MN: A Spatial Modeling Approach

This article examines the social equity of e-scooter availability across differentsocially disadvantaged groups in Minneapolis, MN (USA). The city’snorthwestern part with high poverty rates has a higher e-scooter availability ratethan its counterparts. Also, this study did not find any significant inequalityinvolved with race or educational attainment. However, while the city performedwell on these, areas with a high percentage of commuters, dependent on transit,bike, and walking, have fewer e-scooters available on their streets. As such an area,the local models of this study underscore the need for more e-scooter deploymentin areas close to downtown.