Regression Loss in Transformer-based Supervised Neural Machine Translation

Transformer-based model has achieved human-level performance in supervised neural machine translation (SNMT), much better than the models based on recurrent neural networks (RNNs) or convolutional neural network (CNN). The original Transformer-based model is trained through maximum likelihood estimation (MLE), which regards the machine translation task as a multilabel classification problem and takes the sum of the cross entropy loss of all the target tokens as the loss function. However, this model assumes that token generation is partially independent, without realizing that tokens are the components of a sequence. To solve the problem, this paper proposes a semantic regression loss for Transformer training, treating the generated sequence as a global. Upon finding that the semantic difference is proportional to candidate-reference distance, the authors considered the machine translation problem as a multi-task problem, and took the linear combination of cross entropy loss and semantic regression loss as the overall loss function. The semantic regression loss was proved to significantly enhance SNMT performance, with a slight reduction in convergence speed.

Radio Propagation Measurements in the Indoor Stairwell Environment at 3.5 and 28 GHz for 5G Wireless Networks

To cover the high demand for wireless data services for different applications in the wireless networks, different frequency bands below 6 GHz and in millimeter-wave (mm-Wave) above 24 GHz are proposed for the fifth generation (5G) of communication. The communication network is supposed to handle, among others, indoor traffic in normal situations as well as during emergencies. The stairway is one of those areas which has less network traffic during normal conditions but increases significantly during emergencies. This paper presents the radio propagation in an indoor stairway environment based on wideband measurements in the line of sight (LOS) at two candidate frequencies for 5G wireless networks, namely, 3.5 GHz and 28 GHz. The path loss, root mean square (RMS) delay spread, K-factor results, and analysis are provided. The close-in free-space reference distance (CI), floating intercept (FI), and the close-in free-space reference distance with frequency weighting (CIF) path loss models are provided. The channel parameters such as the number of clusters, the ray and cluster arrival rates, and the ray and cluster decay factors are also obtained for both frequencies. The findings of the path loss show that the CI, FI, and CIF models fit the measured data well in both frequencies with the path loss exponent identical to the free-space path loss. Based on clustering results, it is found that the cluster decay rates are identical at both bands. The results from this and previous measurements indicate that at least one access point is required for every two sections of the stairway to support good coverage along the stairwell area in 5G wireless networks. Moreover, for 5G systems utilizing mm-Wave frequency bands, one access point for each stair section might be necessary for increased reliability of the 5G network in stairwell environments.

Mobile 4G LTE Networks Mobility and Coverage for Some Locations in Ibadan Using Path Loss Analysis

The recent migration of most GSM service providers’ networks in Ibadan from 3G to 4G, in preparation for the deployment of 5G technology in the nation necessitated the need to re-examine the GSM networks’ mobility and coverage within the micro cells in-between a Base Station and a Mobile Station. This attempt is aimed at using existing Path Loss Propagation Models in proffering solutions to the negative consequences usually associated with call drops in the Urban and Suburban Areas of Ibadan due to inability of channels to handover as a result of path loss. The path loss (dB) analysis was carried out by measuring the Received Signal Strength RSS (dBm) at distances ranging from 0.05 km to 4 km in-between Base Stations and Mobile Stations using the factory fitted installed RSS software on Android phones. These measurements were taken for three weeks within the urban and suburban areas of the University of Ibadan campus, and its neighborhood community of Agbowo for ten selected existing Base Stations from four of the nationwide GSM Service Providers (SP1, SP2, SP3 and SP4) in Nigeria. The variation of path loss with the RSS for GSM Service Provider (SP1) propagating at 955MHz (reference distance of 0.05km), 1850MHz and 2120MHz, were 66.03 dB, 71.77 dB and 72.96 dB, respectively. However, at 4 km the path loss had risen to 101.59dB for 955MHz, 103.81dB for 1850MHz and 105dB for 2120MHz. Also, the path loss for the GSM service provider (SP2) propagating at 960MHz (reference distance of 0.05km), 1865MHz and 2150MHz were 66.07 dB, 71.84 dB and 73.08 dB, respectively. Moreover, in a similar manner to the SP1 service provider, at 4 km the path loss had risen to 104.14dB for 960MHz, 109.9dB for 1865MHz and 111.14dB for 2150MHz. Furthermore, the path loss for the GSM service provider (SP3) propagating at 950MHz (reference distance of 0.05km), 1835MHz and 2130MHz were 65.98 dB, 71.70 dB and 73.00 dB, respectively. Likewise, as was in the case of the SP1 and SP2 Service providers, the path loss at 4 km had risen to 104.05dB for 950MHz, 109.76dB for 1835MHz and 111.06dB for 2130MHz. Also, the path loss for the GSM service provider (SP4) propagating at 940MHz (reference distance of 0.05km), 1880MHz and 2140MHz, were 65.47 dB, 71.46 dB and 72.23 dB, respectively. Moreover, the path loss at 4 km had risen to 103.53dB for 940MHz, 109.52dB for 1880MHz and 110.29dB for 2140 MHz as was the case with the other GSM Service providers (SP!, SP2 and SP3) considered .Thus, the path loss increases with distance within the microcells of base stations. However, the path loss model with minimum path loss (dB) at a given distance enhances good coverage and handover postponement. Moreover, the mean square error values used in obtaining the accuracy between the measured and the Empirical models were 17.15dB, 59.69dB, 48.46dB, 60.52dB and 40.07dB for the C-model, Cost-OH, Sub-O, Lee-model and experimental model, respectively. . Key words: GSM networks, Base station, Mobile station, Signal strength, GSM service provider

Path Loss Characterization in an Indoor Laboratory Environment at 3.7 GHz in in Line-Of-Sight Condition

The objective of this work is to propose experimental path loss propagation models for communication channels in indoor environments. In this sense, an experimental path loss characterization has been achieved, according to the measurements campaign carried out in a typical scenario of a university campus. These narrowband measurements were collected in the laboratory environment at 3.7 GHz in line-of-sight (LOS) condition. Also, these measurements were carried out at night to simulate stationary channel conditions. Thus, the results obtained show the values of the parameters of the close-in (CI) free space reference distance and floating-intercept (FI) path loss models, in terms of the transmitter and receiver separation distance. It should be noted that these values of the path loss models have been extracted applying linear regression techniques to the measured data. Also, these values agree with the path loss exponent values presented by other researchers in similar scenarios. The path loss behavior can be described with the implementation of these models. However, more measurement campaigns are needed to improve the understanding of propagation channel features, as well as to obtain better precision in the results obtained. This, in order to optimize the deployment and performance of next fifth-generation (5G) networks that combine indoor environments to offer their services and applications.

Audio-visual interactions in egocentric distance perception: Ventriloquism effect and aftereffect

The processes of audio-visual integration and of visually-guided re-calibration of auditory distance perception are not well understood. Here, the ventriloquism effect (VE) and aftereffect (VAE) were used to study these processes in a real reverberant environment. Auditory and audio-visual (AV) stimuli were presented, in interleaved trials, over a range of distances from 0.7 to 2.04 m in front of the listener, whose task was to judge the distance of auditory stimuli or of the auditory components of AV stimuli. The relative location of the visual and auditory components of AV stimuli was fixed within a session such that the visual component was presented from distance 30% closer (V-closer) than the auditory component, 30% farther (V-farther), or aligned (V-aligned). The study examined the strength of VE and VAE as a function of the reference distance and of the direction of the visual component displacement, and the temporal profile of the build-up/break-down of these effects. All observed effects were approximately independent of target distance when expressed in logarithmic units. The VE strength, measured in the AV trials, was roughly constant for both directions of visual-component displacement such that, on average, the responses shifted in the direction of the visual component by 72% of the audio-visual disparity. The VAE strength, measured on the interleaved auditory-only trials, was stronger in the V-farther than the V-closer condition (44% vs. 31% of the audio-visual disparity, respectively). The VAE persisted to post-adaptation auditory-only blocks of trials, however it was weaker and the V-farther/V-closer asymmetry was reduced. The rates of build-up/break-down of the VAE were also asymmetrical, with slower adaptation in the V-closer condition. These results suggest that, on a logarithmic scale, the AV distance integration is symmetrical, independent of the direction of induced shift, while the visually-induced auditory distance re-callibration is asymmetrical, stronger and faster when evoked by more distant visual stimuli.

Biosonar spatial resolution along the distance axis: revisiting the clutter interference zone

Unlike all other remote senses like vision or hearing, echolocation allows estimating the distance of an object. Not only have echolocating bats and toothed whales been shown to measure distance by echolocation extremely precisely, distance information is even topographically represented by a neuro-computational map in bats’ auditory cortex. This topographic representation and the corresponding tuning of cortical cells to object distance suggests the bats may be able to perceptually resolve multiple, simultaneously present objects along the distance axis. Here we use a novel psychophysical paradigm with complex phantom targets to quantity spatial resolution along the distance axis in the echolocating bat Phyllostomus discolor. We show that our bats can indeed perceptually resolve objects along the distance axis when they are separated by about 40 cm (around a reference distance of 108 cm) along the distance axis. These results are well comparable to earlier work on bats’ clutter interference zone (Simmons et al., 1988) and confirm those results with a more robust psychophysical paradigm.Summary statementEcholocating bats perceive absolute distance to objects by measuring the time delay between call and echo. In addition, they possess spatial resolution along the distance axis.

PERFORMANCE ANALYSIS OF 5G PATH LOSS MODELS FOR RURAL MACROCELL ENVIRONMENT

5G networks are expected to use the Millimeter Wave (mmWave) frequency band and this frequency provides wider bandwidth allowing a better quality of service to be offered to the users. However, the mmWave frequencies may lead to a higher path loss due to several factors including blockages,rain and atmosphere. Therefore, to allow optimal positioning of the 5G base stations, the study of path loss model in this 5G mmWave frequencies is crucial. This paper investigates the 5G path loss models as well as their parameters that are most suitable for cross-polarized antennas under rural macrocell environment in Malaysia. Path loss models namely Close In Free Space Reference Distance Path Loss Model (CI) model, and Alpha Beta Gamma (ABG) or Floating Intercept (FI) Model along with their parameters achieved from the previous studies were evaluated by comparing the parameters and models that are closest to the sampled path loss when using antennas that have different patterns and polarizations in an open-source simulator. Results obtained indicate that FI model can be adapted to the majority of the environment where this model showed the lowest Root Mean Square Error (RMSE). The study of path loss models by using advanced simulator or field measurement, and studies on other rural areas from other states in Malaysia will be considered in future works.

Perspective construal techniques in indirect non-referential naming

The research explores the perspective construal techniques applied in predicate indirect noun groups in Russian. In this case, the discourse perspective is construed with a highly salient ob-ject of perspective in the construed frame of reference. To achieve this effect, the speaker / narrator chooses a particular type of predicate indirect noun groups, such as predicate con-structions with the verbs of fictive motion, appearance, and being (existence) or comparative constructions. Each of these construction types demonstrates its own linguistic and cognitive features, which are used to apply various perspectivization techniques to ensure that its inter-pretation will proceed successfully. To detect these techniques, a complex procedure is applied. Hand-selected fragments analysis followed by corpus statistic and correlation analysis help define the parameters and values in predicate indirect noun group constructions. These parameters are referential (bodily modus type, referent type, referent focus type), lexical (first / repeated lexeme use, type of attribute in pre-position, intensifier type in pre-position), syntactic (sentence initial / final position, position before a clause, co-reference distance in words and propositions) and textual ones (textual role, new microevent introduction). Vari-ance analysis has revealed a group of parameters typical for the studied construction types of predicate indirect noun groups. Parametric results allow describe the typical techniques of object mental scanning, object construal, frames of reference (coordinate system) construal. These include the techniques of mental path shortening / prolongation, embodiment con- strual alleviation / constraining, animated / non-animated object construal, stability / instabil-ity of frames of reference, etc.