Search

In a world in which the internet gives us access to any kind of information, there are still limitations when the source of such information is presented in another language. Online translators are far from perfect, this is why language machine translation is a trending topic in the field of deep learning. The purpose of this project is to use the Transformer architecture, developed by Google in 2017, in the context of Multilingual Machine Translation and to improve its results both in translation and a common intermediate representation. The Transformer model is focused on self-attention and composed by an encoder and decoder that rely on a common intermediate representation of the source language. For the purpose of raising the BLEU score that defines the quality of the translation and enhancing the common intermediate representation, we have introduced Part-Of-Speech tagging in the encoder of the model. We perform experiments with four languages (English, Spanish, French and German) both in Machine Translation and in Cross-lingual Natural Language Inference. Finally, we visualize the intermediate representation and make experiments to see how source embeddings codify gender information. Comparing a baseline model without tagging with the new POS tagged codes, the translation BLEU has decreased 0.50 points on average. In the case of NLI, the accuracies have also decreased 8% on average, showing that the POS tagged models do not improve the performances of these tasks. However, in the gender experiments of the encoder embeddings, the accuracy of the gender classification for professions has increased by 1.1%. En un mundo donde Internet nos da acceso a cualquier tipo de información, aún hay limitaciones cuando la fuente de esta se presenta en otro idioma. Los traductores en línea están lejos de ser perfectos y es por eso que la traducción automática entre idiomas es un tema de tendencia en el campo del aprendizaje profundo. El objetivo de este proyecto es utilizar la arquitectura del Transformer, desarrollada por Google en 2017, en el contexto de la traducción automática multilingüe. El modelo se centra en el concepto de la atención y está compuesto por un codificador y un decodificador basados en una representación intermedia del lenguaje de origen. Con el objetivo de aumentar la BLEU, puntuación que define la calidad de la traducción, y mejorar la representación intermedia, hemos introducido el etiquetado de categorías gramaticales (Part-Of-Speech) en el codificador del modelo. Hemos realizado experimentos con cuatro idiomas (inglés, español, francés y alemán) tanto en traducción automática como en Comprensión del Lenguaje Natural (CLN). Finalmente, hemos visualizado la representación intermedia y hemos hecho experimentos para ver cómo las representaciones de palabras en vectores codifican la información de género. Al comparar un modelo base sin etiquetado Part-Of-Speech con los nuevos códigos etiquetados, la BLEU de traducción ha disminuido 0.50 puntos de media. En el caso de la CLN, las precisiones también han disminuido un 8%, demostrando que los modelos etiquetados no mejoran el rendimiento de estas tareas. Sin embargo, la precisión de la clasificación de género para profesiones ha aumentado en un 1.1%. En un món on l'Internet ens dóna accés a qualsevol tipus d'informació, encara hi ha limitacions quan la font d'aquesta informació es presenta en un altre idioma. Els traductors en línia són lluny de ser perfectes i és per això que la traducció automàtica entre idiomes és un tema de tendència en el camp de l'aprenentatge profund. L'objectiu d'aquest projecte és utilitzar l'arquitectura del Transformer, desenvolupada per Google al 2017, en el context de la traducció automàtica multilingüe. El model es centra en el concepte de l'atenció i està compost per un codificador i un decodificador basats en una representació intermèdia del llenguatge d'origen. Amb l'objectiu d'augmentar la BLEU, puntuació que defineix la qualitat de la traducció, i millorar la representació intermèdia comuna, hem introduït l'etiquetatge de categories gramaticals (Part-Of-Speech) al codificador del model. Hem realitzat experiments amb quatre idiomes (anglès, castellà, francès i alemany) tant en traducció automàtica com en Comprensió del Llenguatge Natural (CLN). Finalment, hem visualitzat la representació intermèdia i hem fet experiments per veure com les representacions de paraules en vectors codifiquen la informació de gènere. En comparar un model base sense etiquetatge Part-Of-Speech amb els nous codis etiquetats, la BLEU de traducció ha disminuït 0.50 punts de mitjana. En el cas de la CLN, les precisions també han disminuït un 8%, demostrant que els models etiquetats no milloren el rendiment d'aquestes tasques. No obstant això, la precisió de la classificació de gènere per a professions ha augmentat en un 1.1%.

Despite significant progress on current state-of-the-art image generation models, synthesis of document images containing multiple and complex object layouts is a challenging task. This paper presents a novel approach, called DocSynth, to automatically synthesize document images based on a given layout. In this work, given a spatial layout (bounding boxes with object categories) as a reference by the user, our proposed DocSynth model learns to generate a set of realistic document images consistent with the defined layout. Also, this framework has been adapted to this work as a superior baseline model for creating synthetic document image datasets for augmenting real data during training for document layout analysis tasks. Different sets of learning objectives have been also used to improve the model performance. Quantitatively, we also compare the generated results of our model with real data using standard evaluation metrics. The results highlight that our model can successfully generate realistic and diverse document images with multiple objects. We also present a comprehensive qualitative analysis summary of the different scopes of synthetic image generation tasks. Lastly, to our knowledge this is the first work of its kind.

This paper studies household spending responses to anticipated changes in the consumption tax. To do so, I construct a life-cycle heterogeneous-agent general equilibrium model with durables. The model features a wedge in durable transactions that reflects the actual consumption tax system: households pay the tax when buying the durables but do not receive the tax when selling them. There are three main findings. First, the baseline model reproduces an empirically consistent dynamic pattern of tax elasticity of the taxable spendings. Second, I find that life-cycle is a key component to match the level of tax elasticity of durable spending. Third, the baseline model generates smaller stockpiling of durables based on realistic motive than a model without the wedge. I then use the model for two counter-factual experiments.The first counter-factual experiment finds that the effect of a consumption tax cut is not symmetric to the tax hike. The second counter-factual experiment which compares a one-time tax hike and a multiple-times tax hike shows the multiple-times tax hike scheme generates smaller welfare cost than one-time tax hike.

We are interested in the regression of data to a parameterised system of differential equations formalising a dynamical system. We study the case of the Green Solow model, a neoclassical economics model for sustainable growth. Faced with the challenges posed by the coupling of the equations, the scarcity of data, and their auto-correlation, we devise several solutions. We present a baseline model and propose three models leveraging neural ordinary differential equations, a recently proposed machine learning model. We empirically and comparatively evaluate the performance of the four models. The results demonstrate the advantages of the proposed approach using neural ordinary differential equations. We conclude by discussing the generality of this knowledge- and data-driven approach to the analysis of dynamical systems.

This chapter presents a dynamic analysis on a laminated rubber–metal spring (LR-MS) using an analytical method. Rubber material was chosen because it has very high damping, and so is highly suitable as a vibration isolator application. A rubber-based vibration isolator also offers better environmental sustainability performance due to its natural origin which makes its renewable, recyclable, and biodegradable. Using an analytical method, the internal resonance equation was derived by using the wave propagation method in a longitudinal direction. Then, the lumped parameter system was developed to represent the baseline model of LR-MS. A nondispersive finite rod model has been developed using impedance matrix, and later, wave and mass effects were investigated. The mathematical modeling of the model was developed based on the internal resonance, lumped parameter, and finite rod model. In conclusion, the developed LR-MS mathematical model can be used for future sustainable design of a vibration isolator in building applications.

Robustness against variations in color and resolution of digitized whole-slide images (WSIs) is an essential requirement for any computer-aided analysis in digital pathology. One common approach to encounter a lack of heterogeneity in the training data is data augmentation. We investigate the impact of different augmentation techniques for whole-slide cartography in colon cancer histology using a newly created multi-scanner database of 39 slides each digitized with six different scanners. A state of the art convolutional neural network (CNN) is trained to differentiate seven tissue classes. Applying a model trained on one scanner to WSIs acquired with a different scanner results in a significant decrease in classification accuracy. Our results show that the impact of resolution variations is less than of color variations: the accuracy of the baseline model trained without any augmentation at all is 73% for WSIs with similar color but different resolution against 35% for WSIs with similar resolution but color deviations. The grayscale model shows comparatively robust results and evades the problem of color variation. A combination of multiple color augmentations methods lead to a significant overall improvement (between 33 and 54 percentage points). Moreover, fine-tuning a pre-trained network using a small amount of annotated data from new scanners benefits the performance for these particular scanners, but this effect does not generalize to other unseen scanners.

This paper documents the UBC Linguistics team’s approach to the SIGMORPHON 2021 Grapheme-to-Phoneme Shared Task, concentrating on the low-resource setting. Our systems expand the baseline model with simple modifications informed by syllable structure and error analysis. In-depth investigation of test-set predictions shows that our best model rectifies a significant number of mistakes compared to the baseline prediction, besting all other submissions. Our results validate the view that careful error analysis in conjunction with linguistic knowledge can lead to more effective computational modeling.

Software development effort estimation (SDEE) generally involves leveraging the information about the effort spent in developing similar software in the past. Most organizations do not have access to sufficient and reliable forms of such data from past projects. As such, the existing SDEE methods suffer from low usage and accuracy. We propose an efficient SDEE method for open source software, which provides accurate and fast effort estimates. The significant contributions of our paper are i) Novel SDEE software metrics derived from developer activity information of various software repositories, ii) SDEE dataset comprising the SDEE metrics' values derived from $\approx13,000$ GitHub repositories from 150 different software categories, iii) an effort estimation tool based on SDEE metrics and a software description similarity model. Our software description similarity model is basically a machine learning model trained using the Paragraph Vectors algorithm on the software product descriptions of GitHub repositories. Given the software description of a newly-envisioned software, our tool yields an effort estimate for developing it. Our method achieves the highest Standard Accuracy score of 87.26% (with cliff's $\delta$=0.88 at 99.999% confidence level) and 42.7% with the Automatic Transformed Linear Baseline model. Our software artifacts are available at https://doi.org/10.5281/zenodo.5095723., Comment: 45 pages, 10 figures, 11 tables, 3 algorithms, Accepted in ACM TOSEM

Path prediction of autonomous vehicles is an essential requirement under any given traffic scenario. Trajectory of several agent vehicles in the vicinity of ego vehicle, at least for a short future, is needed to be predicted in order to decide upon the maneuver of the ego vehicle. We explore variational autoencoder networks to obtain multimodal trajectories of agent vehicles. In our work, we condition the network on past trajectories of agents and traffic scenes as well. The latent space representation of traffic scenes is achieved by using another variational autoencoder network. The performance of the proposed networks is compared against a residual baseline model.

Multimodal summarization becomes increasingly significant as it is the basis for question answering, Web search, and many other downstream tasks. However, its learning materials have been lacking a holistic organization by integrating resources from various modalities, thereby lagging behind the research progress of this field. In this study, we release a full-scale multimodal dataset comprehensively gathering documents, summaries, images, captions, videos, audios, transcripts, and titles in English from CNN and Daily Mail. To our best knowledge, this is the first collection that spans all modalities and nearly comprises all types of materials available in this community. In addition, we devise a baseline model based on the novel dataset, which employs a newly proposed Jump-Attention mechanism based on transcripts. The experimental results validate the important assistance role of the external information for multimodal summarization.

Motorcycles are often related to transit and criminal issues due to its abundance in the transit. Despite its importance, motorcycles are a seldom addressed problem in the computer vision community. We credit this problem to the lack of large-scale datasets and strong baseline models. Therefore, we present the first large-scale Motorcycles Re-Identification (MoRe) dataset. MoRe consists of 3,827 individuals (i.e., the set of motorbikes and motorcyclist) captured by ten surveillance cameras placed in Brazil’s urban traffic scenarios. Furthermore, we evaluate a deep learning model trained using well-known training tricks from the object re-identification literature to present a strong baseline for the motorcycle re-identification (ReID) problem. More importantly, we highlight some crucial problems in this topic as the influence of distractors and the domain shift. Experimental results demonstrate the effectiveness of the strong baseline model with an increase of at least 19.27 p.p. in the rank-1 when compared to the state-of-the-art in the BPReID dataset. Finally, we present some insights regarding the information learned by the strong baseline model when computing the similarities between motorcycle images.1

Background and Objectives: Alloimmunization is a well-known adverse event associated with red blood cell (RBC) transfusions, caused by phenotype incompatibilities between donor and patient RBCs that may lead to haemolytic transfusion reactions on subsequent transfusions. Alloimmunization can be prevented by transfusing fully matched RBC units. Advances in RBC genotyping render the extensive typing of both donors and patients affordable in the foreseeable future. However, the exponential increase in the variety of extensively typed RBCs asks for a software-driven selection to determine the ‘best product for a given patient’. Materials and Methods: We propose the MINimize Relative Alloimmunization Risks (MINRAR) model for matching extensively typed RBC units to extensively typed patients to minimize the risk of alloimmunization. The key idea behind this model is to use antigen immunogenicity to represent the clinical implication of a mismatch. Using simulations of non-elective transfusions in Caucasian donor and patient populations, the effect on the alloimmunization rate of the MINRAR model is compared with that of a baseline model that matches antigens A, B and RhD only. Results: Our simulations show that with the MINRAR model, even for small inventories, the expected number of alloimmunizations can be reduced by 78.3% compared with a policy of only matching on antigens A, B and RhD. Furthermore, a reduction of 93.7% can be achieved when blood is issued from larger inventories. Conclusion: Despite an exponential increase in phenotype variety, matching of extensively typed RBCs can be effectively implemented using our MINRAR model, effectuating a substantial reduction in alloimmunization risk without introducing additional outdating or shortages.

In this paper we present SentiTel, a fine-grained opinion mining dataset that is human annotated for the task of targeted aspect-based sentiment analysis (TABSA). SentiTel contains Twitter reviews about three telecoms in Uganda posted in the period between February 2019 and September 2019. The reviews in the dataset have a code-mix of English and Luganda a language that is commonly spoken in Uganda. The dataset in this paper consists of 5973 human annotated reviews with the target entity which is the target telecom, aspect and sentiment towards the aspect of the target telecom. Each review contains at least one target telecom. Two models are trained for the TABSA task that is random forest which is the baseline model and BERT. The best results are obtained using BERT with an Area Under the ROC Curve (AUC) of 0.950 and 0.965 on aspect category detection and sentiment classification respectively. The results show that even though tweets are written without the intention of writing a formal review, they are rich in information and can be used for fine-grained opinion mining. Finally, the results show that fine-tuning the pre-trained BERT model on a downstream task generates better results compared to the baseline models.

Coronavirus disease 2019 (COVID-19), the pandemic that is spreading fast globally, has caused over 34 million confirmed cases. Apart from the reverse transcription polymerase chain reaction (RT-PCR), the chest computed tomography (CT) is viewed as a standard and effective tool for disease diagnosis and progression monitoring. We propose a diagnosis and prognosis model based on graph convolutional networks (GCNs). The chest CT scan of a patient, typically involving hundreds of sectional images in sequential order, is formulated as a densely connected weighted graph. A novel distance aware pooling is proposed to abstract the node information hierarchically, which is robust and efficient for such densely connected graphs. Our method, combining GCNs and distance aware pooling, can integrate the information from all slices in the chest CT scans for optimal decision making, which leads to the state-of-the-art accuracy in the COVID-19 diagnosis and prognosis. With less than 1% number of total parameters in the baseline 3D ResNet model, our method achieves 94.8% accuracy for diagnosis. It has a 2.4% improvement compared with the baseline model on the same dataset. In addition, we can localize the most informative slices with disease lesions for COVID-19 within a large sequence of chest CT images. The proposed model can produce visual explanations for the diagnosis and prognosis, making the decision more transparent and explainable, while RT-PCR only leads to the test result with no prognosis information. The prognosis analysis can help hospitals or clinical centers designate medical resources more efficiently and better support clinicians to determine the proper clinical treatment.

Utilizing clinical texts in survival analysis is difficult because they are largely unstructured. Current automatic extraction models fail to capture textual information comprehensively since their labels are limited in scope. Furthermore, they typically require a large amount of data and high-quality expert annotations for training. In this work, we present a novel method of using BERT-based hidden layer representations of clinical texts as covariates for proportional hazards models to predict patient survival outcomes. We show that hidden layers yield notably more accurate predictions than predefined features, outperforming the previous baseline model by 5.7% on average across C-index and time-dependent AUC. We make our work publicly available at https://github.com/bionlplab/heart_failure_mortality., Comment: NAACL-HLT 2021, Short Paper

Automatic co-text free name matching has a variety of important real-world applications, ranging from fiscal compliance to border control. Name matching systems use a variety of engines to compare two names for similarity, with one of the most critical being phonetic name similarity. In this work, we re-frame existing work on neural sequence-to-sequence transliteration such that it can be applied to name matching. Subsequently, for performance reasons, we then build upon this work to utilize an alternative, non-recurrent neural encoder module. This ultimately yields a model which is 63% faster while still maintaining a 16% improvement in averaged precision over our baseline model.

Enabling machines to read, process, and understand natural language documents is a coveted goal of artificial intelligence. However, this task is extremely challenging, and most existing models lack the ability to perform complex reasoning. Considering that humans often read documents roughly first when understanding a problem, this paper proposes a new model that attempts to mimic the reasoning process of human readers. Our model performs a extensive read and a intensive read of the document separately, and then combines the information obtained from both reading methods to finally find a satisfactory answer. Finally, by experimenting within RACE dataset and comparing with the baseline model BERT, the feasibility and effectiveness of our proposed model can be illustrated.

International audience; Automatic generation of advertising layouts shows high economic interest, but as identified with our industrial partner, there is no public document layout dataset that matches this particular application. In this context, we produced two synthetic datasets that allow both the evaluation and training of any learning model on web advertising layout generation, and a small dataset of real cases to demonstrate the contribution of our work. We compared the results obtained by different learning models on the real cases, with and without prior use of our synthetic datasets, and our results show that these datasets allow to build and decisively improve models for the generation of real-world advertising layouts. Our three datasets, as well as useful data processing tools, are

Despite the benefits associated with the free movement of people, governments often try to regulate urban immigration by constraining the agency of potential rural out-migrants in moving to cities and/or in expanding their agency to enable them to stay put. We apply an institutional framework centring on push–pull and retain–repel factors to migration intentions of potential migrants in northern Kazakhstan. We model the effects of these factors on migration intentions with Bayesian Networks and expand the baseline model with three policy scenarios. The results suggest that the effects of policies constraining urban in-migration, e.g. limiting access to affordable housing, are attenuated by social networks and reverse remittances. The supply of accessible and appropriate information on possible income and true housing costs in urban areas presents a promising road to reduce intentions of rural out-migration. Better schools and decentralised tertiary education can also reduce the migration intentions of rural residents. En dépit des avantages associés à la libre circulation des personnes, les gouvernements tentent souvent de réguler l'immigration urbaine en limitant la capacité de potentiels émigrants ruraux à se déplacer vers les villes et / ou en élargissant leur capacité à rester sur place. Nous appliquons un cadre institutionnel centré sur les facteurs push-pull et les facteurs de rétention-répulsion liés aux intentions de migration de potentiels migrants dans le nord du Kazakhstan. Nous modélisons grâce aux réseaux bayésiens les effets de ces facteurs sur les intentions de migration et élargissons le modèle de référence avec trois scénarios de politiques publiques. Les résultats suggèrent que l’effet des politiques publiques limitant l'immigration urbaine, comme par exemple limiter l'accès à des logements à loyer abordable, est atténué par les réseaux sociaux et par les transferts de fonds inversés. Le fait de fournir des informations accessibles et pertinentes sur ce que l’on peut gagner, en termes de revenus, et sur ce que l’on doit dépenser pour se loger dans les zones urbaines représente une voie prometteuse pour réduire les intentions d'exode rural. De meilleures écoles et un enseignement supérieur décentralisé peuvent également réduire les intentions de migration des résidents ruraux.

Text generation is one of the most important tasks in NLP and has been applied in many applications such as machine translation, question answering and text summarization. Most of recent studies on text generation use only the input for output generation. In this research we suggest that topic information of an input document is an important factor for generating the destination text. We will propose a deep neural network model in which we use topic information together with the input text for generating summarized texts. The experiment on Vietnamese news corpus shows that our model outperforms a baseline model at least 23% in BLEU score.

The large-scale shifts in weather patterns and an unprecedented change in climate have given rise to the interest in how climate change will affect the carbon emissions of supermarkets. This study investigates the implications of future climatic conditions on the operation of supermarkets in the UK. The investigation was conducted by performing a series of energy modelling simulations on a LIDL supermarket model in London, based on the UK Climate Projections (UKCP09) future weather years provided by the Chartered Institution of Building Services Engineers (CIBSE). Computational fluid dynamic (CFD) simulations were used to perform the experiment, and the baseline model was validated against the actual data. This investigation ascertains and quantifies the annual energy consumption, carbon emissions, and cooling and heating demand of the supermarket under different climatic projections, which further validate the scientific theory of annual temperature rise as a result of long-term climatic variation. The maximum percentage increase for the annual energy consumption for current and future weather data sets observed was 7.01 and 6.45 for the 2050s medium emissions scenario, (90th) percentile and high emissions scenario, (90th) percentile, respectively, and 11.05, 14.07, and 17.68 for the 2080s low emissions scenario, (90th) percentile, medium (90th) percentile and high emissions scenario (90th) percentile, respectively. A similar inclining trend in the case of annual CO2 emissions was observed where the peak increase percentage was 6.80 and 6.24 for the 2050s medium emissions scenario, (90th) percentile and high (90th) percentile, respectively and 10.84, 13.84, and 17.45 for the 2080s low emissions scenario, (90th) percentile, medium emissions scenario (90th) percentile and high emissions scenario (90th) percentile, respectively. The study also analyses the future heating and cooling demands of the three warmest months and three coldest months of the year, respectively, to determine future variance in their relative values.

The last two decades have seen an exponential increase in the use of the Internet and social media, which has changed basic human interaction. This has led to many positive outcomes. At the same time, it has brought risks and harms. The volume of harmful content online, such as hate speech, is not manageable by humans. The interest in the academic community to investigate automated means for hate speech detection has increased. In this study, we analyse six publicly available datasets by combining them into a single homogeneous dataset. Having classified them into three classes, abusive, hateful or neither, we create a baseline model and improve model performance scores using various optimisation techniques. After attaining a competitive performance score, we create a tool that identifies and scores a page with an effective metric in near-real-time and uses the same feedback to re-train our model. We prove the competitive performance of our multilingual model in two languages, English and Hindi. This leads to comparable or superior performance to most monolingual models.

Infants cry for many different reasons. Understanding the infant’s language is a critical challenge that many parents suffer from, thus, it is hard to know precisely for what reason infants are crying. The purpose of our study is to determine whether the infant cry is due to hunger or not, using semi-supervised machine learning techniques. There are two commonly used datasets in the literature, the Dunstan Baby Language and Baby Chillanto database. The total length of each of the datasets is only between 8 and 32 minutes, which is very short. For this reason, we proposed a semi-supervised learning approach (also known as self-training), which can increase the dataset by classifying the unlabeled data from Google AudioSet. We have chosen the k-nearest neighbors (KNN) classifier to determine whether the cry is due to hunger or not. The KNN is known to produce low-performance results if trained with limited data. Thus, we proposed our semi-supervised k-nearest neighbor (SSKNN) that can benefit from unlabeled data to increase the training set. As for feature extraction, we chose Mel Frequency Cepstral Coefficient. To evaluate the performance of the semi-supervised approach, we used the supervised KNN as our baseline model and compared the accuracy between the two approaches. The SSKNN yields better accuracy, which is 94% compared to the supervised KNN which has only an accuracy of 87%.

It is one of the important tasks to predict a participant's role in a multi-party conversation. Many previous studies utilized only verbal or non-verbal features to construct models for the role recognition task. In this paper, we propose a model that combines verbal and non-verbal features for leader identification. We add non-verbal features and construct our prediction model with utterance, pose, facial, and prosodic features. In our experiments, we compare our model with a baseline model that is based on only utterance features. The results show the effectiveness of our multimodal approach. In addition, we improve the performance of the baseline model to add some new utterance features., International Conference on Asian Language Processing (IALP 2020), 4-6 December, 2020, Kuala Lumpur, Malaysia（新型コロナ感染拡大に伴い、オンライン開催に変更）

While reinforcement learning (RL) on humans has shown incredible promise, it often suffers from a scarcity of data and few steps. In instances like these, a planning model of human behavior may greatly help. We present an experimental setup for the development and testing of an Soft Actor Critic (SAC) V2 RL architecture for several different neural architectures for planning models: an autoML optimized LSTM, an OLS, and a baseline model. We present the effects of including a planning model in agent learning within a simulation of the office, currently reporting a limited success with the LSTM.

Derivation of structure-kinetics relationships can help rational design and development of new small-molecule drug candidates with desired residence times. Efforts are now being directed toward the development of efficient computational methods. Currently, there is a lack of solid, high-throughput binding kinetics prediction approaches on bigger datasets. We present a prediction method for binding kinetics based on the machine learning analysis of protein-ligand structural features, which can serve as a baseline for more sophisticated methods utilizing molecular dynamics (MD). We showed that the random forest algorithm is capable of learning the protein binding site secondary structure and backbone/side-chain features to predict the binding kinetics of protein-ligand complexes but still with inferior performance to that of MD-based descriptor analysis. MD simulations had been applied to a limited number of targets and a series of ligands in terms of kinetics analysis, and we believe that the developed approach may guide new studies. The method was trained on a newly curated database of 501 protein-ligand unbinding rate constants, which can also be used for testing and training the binding kinetics prediction models.

Spatiotemporal datasets based on player tracking are widely used in sports analytics research. Common research tasks often require the analysis of game events, such as passes, fouls, tackles, and shots on goal. However, spatiotemporal datasets usually do not include event information, which means it has to be reconstructed automatically. We propose a rule-based algorithm for identifying several basic types of events in soccer, including ball possession, successful and unsuccessful passes, and shots on goal. Our aim is to provide a simple procedure that can be used for practical soccer data analysis tasks, and also serve as a baseline model for algorithms based on more advanced approaches. The resulting algorithm is fast, easy to implement, achieves high accuracy on the datasets available to us, and can be used in similar scenarios without modification.

When only a limited amount of accented speech data is available, to promote multi-accent speech recognition performance, the conventional approach is accent-specific adaptation, which adapts the baseline model to multiple target accents independently. To simplify the adaptation procedure, we explore adapting the baseline model to multiple target accents simultaneously with multi-accent mixed data. Thus, we propose using accent-specific top layer with gate mechanism (AST-G) to realize multi-accent adaptation. Compared with the baseline model and accent-specific adaptation, AST-G achieves 9.8% and 1.9% average relative WER reduction respectively. However, in real-world applications, we can't obtain the accent category label for inference in advance. Therefore, we apply using an accent classifier to predict the accent label. To jointly train the acoustic model and the accent classifier, we propose the multi-task learning with gate mechanism (MTL-G). As the accent label prediction could be inaccurate, it performs worse than the accent-specific adaptation. Yet, in comparison with the baseline model, MTL-G achieves 5.1% average relative WER reduction., Accepted in INTERSPEECH 2019

This paper analyzes the impact of fiscal spending shocks in a multi-country model with international production networks. In contrast to standard results suggesting that production network linkages are unimportant for the aggregate response to macro shocks in a closed economy, we show that network structures may place a central role in the international propagation of fiscal shocks, particularly when wages are slow to adjust. The paper first develops a simple general equilibrium multi-country model and derives some analytical results on the response to fiscal spending shocks. We then apply the model to an analysis of fiscal spillovers in the Eurozone, using the calibrated sectoral network structure from the World Input Output Database (WIOD). In a version of the model with sticky wages, we find that fiscal spillovers from Germany and some other large Eurozone countries may be large, and within the range of empirical estimates. More importantly, we find that the Eurozone production network is very important for the international spillovers. In the absence of international production network linkages, spillovers would be only a third as large as predicted by the baseline model. Finally, we explore the diffusion of identified German government spending at the sectoral level, both within and across countries. We find that government expenditures have both significant upstream and downstream effects when these links are measured by the direction of sectoral production linkages.

Methods relying entirely on visible-range/color images start to have problems in detection tasks when there is not enough light to illuminate the scene. Thermal cameras, which operate based on the infrared radiation emitted by objects, can provide detectable information in low- or no-light conditions. In this paper, we propose a method to improve the performance of a pedestrian detection algorithm on thermal images by incorporating features from saliency maps to enrich the thermal image features. We employ a modified version of a state-of-the-art object detection network, and feed the thermal images and their saliency maps to two parallel networks. Experimental results on five different datasets show that our proposed approach performs better at detecting pedestrians in thermal images compared to its vanilla version and a baseline model.

Research has been conducted on approaches using social media information to improve demand forecasting accuracy in business-to-customer industries. However, such social media information is not applicable to business-to-business (B2B) industries, as a result of a lack of end-consumer evaluations. This raises a few interesting questions, including whether there may be any external information that could be used to improve B2B demand forecasting, and whether practical approaches may be possible to collect and utilize useful external business information. In this study, we develop an innovative and intelligent demand forecasting approach and apply it to a B2B server company based in the United States. We first implemented time series and machine learning models based on sales data and selected the best-fitting model as a baseline, and then used a web crawler and Google Trends to collect related market signals as external information indices for the server industry, which were finally incorporated into the selected baseline model to adjust forecasting results to account for demand fluctuations. Experimental results demonstrate that the baseline model achieved an out‐of‐sample mean squared error (MSE) of 19.77 without considering the collected external information indices, and 11.87 when external information was incorporated. Therefore, our proposed approach significantly improved forecasting accuracy, demonstrating an improvement of 63.1% in terms of MSE, 44.1% in terms of mean absolute error, and 61.2% in terms of root mean square percentage error. Thus, this study sheds light on the value of external information in demand forecasting for B2B industries.

This paper presents an adversarial learning method for recognition-synthesis based non-parallel voice conversion. A recognizer is used to transform acoustic features into linguistic representations while a synthesizer recovers output features from the recognizer outputs together with the speaker identity. By separating the speaker characteristics from the linguistic representations, voice conversion can be achieved by replacing the speaker identity with the target one. In our proposed method, a speaker adversarial loss is adopted in order to obtain speaker-independent linguistic representations using the recognizer. Furthermore, discriminators are introduced and a generative adversarial network (GAN) loss is used to prevent the predicted features from being over-smoothed. For training model parameters, a strategy of pre-training on a multi-speaker dataset and then fine-tuning on the source-target speaker pair is designed. Our method achieved higher similarity than the baseline model that obtained the best performance in Voice Conversion Challenge 2018., Comment: Accepted to INTERSPEECH 2020

Background and objectiveDetecting abnormal patterns within an electrocardiogram (ECG) is crucial for diagnosing cardiovascular diseases. We start from two unresolved problems in applying deep-learning-based ECG classification models to clinical practice: first, although multiple cardiac arrhythmia (CA) types may co-occur in real life, the majority of previous detection methods have focused on one-to-one relationships between ECG and CA type, and second, it has been difficult to explain how neural-network-based CA classifiers make decisions. We hypothesize that fine-tuning attention maps with regard to all possible combinations of ground-truth (GT) labels will improve both the detection and interpretability of co-occurring CAs. Methods To test our hypothesis, we propose an end-to-end convolutional neural network (CNN), xECGNet, that fine-tunes the attention map to resemble the averaged response maps of GT labels. Fine-tuning is achieved by adding to the objective function a regularization loss between the attention map and the reference (averaged) map. Performance is assessed by F1 score and subset accuracy. Results The main experiment demonstrates that fine-tuning alone significantly improves a model’s multilabel subset accuracy from 75.8% to 84.5% when compared with the baseline model. Also, xECGNet shows the highest F1 score of 0.812 and yields a more explainable map that encompasses multiple CA types, when compared to other baseline methods. Conclusions xECGNet has implications in that it tackles the two obstacles for the clinical application of CNN-based CA detection models with a simple solution of adding one additional term to the objective function.

Recent neural text-to-SQL models can effectively translate natural language questions to corresponding SQL queries on unseen databases. Working mostly on the Spider dataset, researchers have proposed increasingly sophisticated solutions to the problem. Contrary to this trend, in this paper we focus on simplifications. We begin by building DuoRAT, a re-implementation of the state-of-the-art RAT-SQL model that unlike RAT-SQL is using only relation-aware or vanilla transformers as the building blocks. We perform several ablation experiments using DuoRAT as the baseline model. Our experiments confirm the usefulness of some techniques and point out the redundancy of others, including structural SQL features and features that link the question with the schema., Accepted to NAACL 2021. 9 pages

This paper proposes a network slimming method for synthetic aperture radar (SAR) ship detection based on knowledge distillation. Firstly, the generic objection detection network is pruned regularly and extremely on filter-level to get lightweight models under different global pruning ratios. Secondly, the knowledge distillation framework based on Kullback Leibler (KL) Divergence is used to train small student network and large teacher network from scratch synchronously to restore the accuracy of student network. To verify the effectiveness of the proposed method, sufficient experiments are conducted on the widely used SAR Ship Detection Dataset (SSDD). YOLO v3@Darknet-53 is selected as the baseline model while YOLO v3@EfficientNet-B7 as the teacher network. Results show that, with our method, a student model with only 15.4M parameters (25% of the baseline model) can achieve high pruning ratio while still maintaining encouraging performance. Compared with the baseline model, there are only 1% and 0.9% differences on average precision (AP) and average recall (AR), respectively. Compared with traditional fine-tuning method which only restores 0.6% AP, the model slimming method based on knowledge distillation proposed in this paper restores 2.4% AP with obvious advantages.

In recent years speech enhancement has shown great progress that was driven mostly by using bigger and more sophisticated neural networks. In this work we investigate the possibility to use state-of-the-art speech enhancement neural network and modify it in such a way that will allow it to process the noisy signal multiple times. By doing so we expect, that with each iteration the enhancement will improve. Experiments conducted using the WSJ0, Noisex-92 and DCASE datasets show, that U-net with gated dilated convolutions is able to achieve better SI-SDR, STOI and PESQ after processing the noisy signal two times, with the improvement being consistent across all SNRs and tested noise types. This is achieved without any additional trainable parameters and no additional memory requirements compared to the baseline model.

Predicting traffic has been important for websites' daily services. Developing efficient models for Wikipedia's page traffic would deepen our knowledge about people's behavior on Wikipedia and potentially for other crowdsourcing pages. The current project attempted to experiment with incorporating time series data from a linked page trying to improve the prediction accuracy of future traffic of a page. The current study experimented with three timeseries models. The baseline model uses the monthly traffic of 2019 of a page to predict the monthly traffic of January of 2020. The random neighbor model randomly selects a page which has a hyperlink to the focal page and uses the 2019 data of the focal page and the neighboring page to predict the monthly traffic of January of 2020. The similar neighbor model also uses data from the focal and a neighboring page, but the neighbor is selected based on its content similarity to the focal page. The results show that prediction with a similar neighbor model has better prediction performance than with the Random neighbor model on popular pages. The baseline model has the best performance with the smallest MSE, MAE, and MAPE, while the random neighbor model and similar neighbor model have much larger MSE than the Baseline model.

In the broader healthcare domain, the prediction bears more value than an explanation considering the cost of delays in its services. There are various risk prediction models for cardiovascular diseases (CVDs) in the literature for early risk assessment. However, the substantial increase in CVDs-related mortality is challenging global health systems, especially in developing countries. This situation allows researchers to improve CVDs prediction models using new features and risk computing methods. This study aims to assess nonclinical features that can be easily available in any healthcare systems, in predicting CVDs using advanced and flexible machine learning (ML) algorithms. A gender-matched case–control study was conducted in the largest public sector cardiac hospital of Pakistan, and the data of 460 subjects were collected. The dataset comprised of eight nonclinical features. Four supervised ML algorithms were used to train and test the models to predict the CVDs status by considering traditional logistic regression (LR) as the baseline model. The models were validated through the train–test split (70:30) and tenfold cross-validation approaches. Random forest (RF), a nonlinear ML algorithm, performed better than other ML algorithms and LR. The area under the curve (AUC) of RF was 0.851 and 0.853 in the train–test split and tenfold cross-validation approach, respectively. The nonclinical features yielded an admissible accuracy (minimum 71%) through the LR and ML models, exhibiting its predictive capability in risk estimation. The satisfactory performance of nonclinical features reveals that these features and flexible computational methodologies can reinforce the existing risk prediction models for better healthcare services.

As fake news and rumors have been known as a mainstream concern, researchers have been finding approaches to identify fake news and rumors on social media platforms. One of the tasks of detecting rumors on social media platforms is RumourEval held by SemEval in 2017 and 2019. This shared task consists of two subtasks where the first task is to classify reactions of a post/comment whereas the second task is rumor veracity, respectfully. Following this topic and using the same dataset provided by the SemEval organizers, we propose a new approach to tackle the problems by augmenting new input data that will be fed into the model, as well as extracting new useful handcrafted Dialog-Act features. Our approach outperforms the baseline model despite using LSTM model.

In this paper, we find a network which can achieve better result than the state of the art for Head and Neck Organ at Risks (OARs) segmentation. At first, we enumerate the popular networks, and sum up their characteristic. We extract the main components from these popular networks and we design experiment to evaluate these components. We split the experiment into two stage. At the first stage experiment, we try to find out which components and constructions can let the network achieve better result than baseline model, Unet, for H&N OAR segmentation. After finding out the useful components and constructions, we try to mix them up to build a novel network which absorbs all their merits. At last, we get a hybrid network, Attention-W-net which gets the best result and defeat the state of the art. All networks are evaluated on 16th CSTRO conference H&N OAR segmentation competition dataset.

Mirnezami, Alexander/0000-0002-6199-8332 WOS:000594274800001 PubMed: 33204304 Imbalanced class distribution in the medical dataset is a challenging task that hinders classifying disease correctly. It emerges when the number of healthy class instances being much larger than the disease class instances. To solve this problem, we proposed undersampling the healthy class instances to improve disease class classification. This model is named Hellinger Distance Undersampling (HDUS). It employs the Hellinger Distance to measure the resemblance between majority class instance and its neighbouring minority class instances to separate classes effectively and boost the discrimination power for each class. An extensive experiment has been conducted on four imbalanced medical datasets using three classifiers to compare HDUS with a baseline model and three state-of-the-art undersampling models. The outcomes display that HDUS can perform better than other models in terms of sensitivity, F1 measure, and balanced accuracy.

The Collaborative Communication Interruption Management System or C-CIMS [1] uses machine learning techniques to build task boundary inference models to send interruptions at appropriate times within distributed multi-user, multitasking interactions. The primary objective of this work is to explore improving C-CIMS performance using speaker and team dependent machine learning techniques. This has the potential to optimize system performance for each talker or team engaged in the interaction. An analysis of variance illustrated that there is a significant difference in C-CIMS performance using the talker-dependent models compared to the team-dependent models. Additionally a subset of talker and teams significantly outperform the baseline model. These results motivate the continued exploration of additional techniques to maximize C-CIMS performance in making improved accurate decisions in disseminating interruptions.

We describe the baseline model configuration and simulation characteristics of the Geophysical Fluid Dynamics Laboratory (GFDL)'s Atmosphere Model version 4.1 (AM4.1), which builds on developments at GFDL over 2013–2018 for coupled carbon‐chemistry‐climate simulation as part of the sixth phase of the Coupled Model Intercomparison Project. In contrast with GFDL's AM4.0 development effort, which focused on physical and aerosol interactions and which is used as the atmospheric component of CM4.0, AM4.1 focuses on comprehensiveness of Earth system interactions. Key features of this model include doubled horizontal resolution of the atmosphere (~200 to ~100 km) with revised dynamics and physics from GFDL's previous‐generation AM3 atmospheric chemistry‐climate model. AM4.1 features improved representation of atmospheric chemical composition, including aerosol and aerosol precursor emissions, key land‐atmosphere interactions, comprehensive land‐atmosphere‐ocean cycling of dust and iron, and interactive ocean‐atmosphere cycling of reactive nitrogen. AM4.1 provides vast improvements in fidelity over AM3, captures most of AM4.0's baseline simulations characteristics, and notably improves on AM4.0 in the representation of aerosols over the Southern Ocean, India, and China—even with its interactive chemistry representation—and in its manifestation of sudden stratospheric warmings in the coldest months. Distributions of reactive nitrogen and sulfur species, carbon monoxide, and ozone are all substantially improved over AM3. Fidelity concerns include degradation of upper atmosphere equatorial winds and of aerosols in some regions.

Indoor Air Quality (IAQ) is one of the utmost concerns when it comes to the health and comfort of the occupants within a structure. However, because of the lack of information with regards to IAQ, especially in the Philippines setting, improvements with respect to the IAQ are not prioritized. Hence, the goal of this study is to determine if the indoor air quality in the Northwest and Southwest buildings of Mapua University is up to standard set by American Society of Heating, Refrigerating and Air-Conditioning Engineers (ASHRAE) and Occupational Safety and Health (OSH). The study was divided into two phases, one focused on the quantitative walkthrough to evaluate IAQ. This was achieved by using an integrated instrument system using thermohygrometers for temperature and relative humidity, Telaire 7001 for CO2 concentration, and SKC IOM Sampler for particulate matter. Results show that in some rooms, temperature exceeded the range for thermal comfort and majority of the rooms does not comply with the ASHRAE standard for CO2 concentration and the particulate matter remained in the same zone in all of the rooms. The second phase of the study focused on the development of predictive models based on the results of the quantitative walkthrough from the first phase. The predictive models were used to predict the ideal number of occupants in a room given the IAQ parameters. Model performance showed that stochastic gradient boosting (gbm) and support vector machine with Radial Basis Function Kernel (svmRadial) are the best performing models with R2 and RMSE of 0.6838, 0.7777 and 0.812, 0.804 respectively.

This paper documents state dependence in labor market \ud uctuations. Using a Threshold\ud Vector Autoregression model (TVAR), we establish that the unemployment rate, the job\ud separation rate, and the job finding rate exhibit a larger response to productivity shocks\ud during periods with low aggregate productivity. A Diamond-Mortensen-Pissarides model\ud with endogenous job separation and on-the-job search replicates these empirical regularities\ud well. We calibrate the model to match the standard deviation of the job-transition rates\ud explained by productivity shocks in the TVAR, and show that the model explains 88 percent\ud of the state dependence in the unemployment rate, 76 percent for the separation rate and\ud 36 percent for the job finding rate. The key channel underpinning state dependence in both\ud job separation and job finding rates is the interaction of the firm's reservation productivity\ud level and the distribution of match-specific idiosyncratic productivity. Results are robust\ud across several variations to the baseline model.