Search

We study pool-based active learning with abstention feedbacks where a labeler can abstain from labeling a queried example with some unknown abstention rate. This is an important problem with many useful applications. We take a Bayesian approach to the problem and develop two new greedy algorithms that learn both the classification problem and the unknown abstention rate at the same time. These are achieved by simply incorporating the estimated average abstention rate into the greedy criteria. We prove that both algorithms have near-optimality guarantees: they respectively achieve a ( 1 - 1 e ) constant factor approximation of the optimal expected or worst-case value of a useful utility function. Our experiments show the algorithms perform well in various practical scenarios.

The problem of monotone missing data has been broadly studied during the last two decades and has many applications in various fields such as bioinformatics or statistics. Commonly used imputation techniques require multiple iterations through the data before yielding convergence. Moreover, those approaches may introduce extra noises and biases to the subsequent modeling. In this work, we derive exact formulas and propose a novel algorithm to compute the maximum likelihood estimators (MLEs) of a multiple class, monotone missing dataset when all the covariance matrices of all categories are assumed to be equal, namely Efficient Parameter Estimation for Multiple Class Monotone Missing Data (EPEM). We then illustrate an application of our proposed methods in Linear Discriminant Analysis (LDA). As the computation is exact, our EPEM algorithm does not require multiple iterations through the data as other imputation approaches, thus promising to handle much less time-consuming than other methods. This effectiveness was validated by empirical results when EPEM reduced the error rates significantly and required a short computation time compared to several imputation-based approaches. We also release all codes and data of our experiments in a GitHub repository to contribute to the research community related to this problem.

The behavior of rPUf (polyurethane foam derived from bis(2-hydroxyethyl) terephthalate (BHET) recovered from post-consumer poly(ethylene terephthalate)-PET bottles) and cPUf (polyurethane foam based on a commercial polyol) were evaluated by comprehensive characterizations, and their thermal stabilities, flammabilities, and physical properties were compared. Without using a flame retardant, rPUf passed the UL-94HB standard, whereas cPUf burned rapidly to the holding clamp. Various formulations of rPUf and cPUf with established eco-friendly fire retardants (FRs), including triphenyl phosphate, aluminum diethylphosphinate, and aluminum trihydroxide, exhibited notably higher fire resistance. The differences between the experimental and calculated densities showed the effectiveness of the foaming process. A larger difference indicated that the FR improved the foaming process more. The BHET structure contains an aromatic moiety that strongly enhanced the compressive strength and protected the rPUf/FRs from biodegradation, as evidenced by sorption isotherm experiments. For that reason, the rPUf/FRs could be used as building materials, especially under high humidity conditions. These are the advantages of BHET over the commercial polyol for PUf preparation. Chemical recycle of post-consumer PET bottles to produce high performance PUf could be applied on an industrial scale and then it creates the sustainable recycling industry.

In this paper, we study the learning rate of generalized Bayes estimators in a general setting where the hypothesis class can be uncountable and have an irregular shape, the loss function can have heavy tails, and the optimal hypothesis may not be unique. We prove that under the multi-scale Bernstein’s condition, the generalized posterior distribution concentrates around the set of optimal hypotheses and the generalized Bayes estimator can achieve fast learning rate. Our results are applied to show that the standard Bayesian linear regression is robust to heavy-tailed distributions.

In 2016, extreme drought and saline intrusion led to widespread damage and livelihood disruption for many people in the coastal region of the Vietnamese Mekong Delta. Through an integrated multi-st...

The objective of this work is to investigate the influence of two phosphorus/phosphorus-nitrogen flame retardants including diammonium hydrogen phosphate (DAP) and aluminum hydrogen phosphite (AHP) on the flame performance of polyurethane/rice husk (PUF-RH) eco-composites. The fire retardancy and thermal stability of the prepared composites were studied by UL-94, LOI, DSC, and TGA. The fire extinguished completely before reaching the 1-inch mark for PUF-RH containing 15 wt% DAP loading and thus the sample achieved the UL-94 HB. Meanwhile, PUR-RH and PUF-RH/AHP15 completely burned and failed the UL-94 HB test. A UL-94 V-0 rating was obtained and LOI value significantly increased by 26.3% at 25 wt% loading of DAP for PUF-RH, which is consistent with enhanced thermal stability confirmed by TGA. DAP in PUF-RH mainly acts in fuel dilution and cooling effect, vapor phase, and condensed phase. Differences in experimental and calculated TG analyses strongly support an interaction between DAP and PUF-RH during their decomposition. The presence of DAP results in the composite with more regularity pore size and a higher density which could be reason for its improved compressive strength. DAP is a non-reactive additive in the polymer system; therefore, it does not influence the Tg of PUF-RH.

Understanding the relationship between objects in an image is an important challenge because it can help to describe actions in the image. In this paper, a graphical data structure, named “Scene Graph”, is utilized to represent an encoded informative visual relationship graph for an image, which we suggest has a wide range of potential applications. This scene graph is applied and tested in the popular domain of lifelogs, and specifically in the challenge of known-item retrieval from lifelogs. In this work, every lifelog image is represented by a scene graph, and at retrieval time, this scene graph is compared with the semantic graph, parsed from a textual query. The result is combined with location or date information to determine the matching items. The experiment shows that this technique can outperform a conventional method.

The Video Browser Showdown (VBS) is an annual competition in which each participant prepares an interactive video retrieval system and partakes in a live comparative evaluation at the annual MMM Conference. In this paper, we introduce Eolas, which is a prototype video/image retrieval system incorporating a novel virtual reality (VR) interface. For VBS’21, Eolas represented each keyframe of the collection by an embedded feature in a latent vector space, into which a query would also be projected to facilitate retrieval within a VR environment. A user could then explore the space and perform one of a number of filter operations to traverse the space and locate the correct result.

There have been a massive number of conferences and journals in computer science that create a lot of difficulties for scientists, especially for early-stage researchers, to find the most suitable venue for their scientific submission. In this paper, we present a novel approach for building a paper submission recommendation system by using two different types of embedding methods, GloVe and FastText, as well as Convolutional Neural Network (CNN) and LSTM to extract useful features for a paper submission recommendation model. We consider seven combinations of initial attributes from a given submission: title, abstract, keywords, title + keyword, title + abstract, keyword + abstract, and title + keyword + abstract. We measure these approaches’ performance on one dataset, presented by Wang et al., in terms of top K accuracy and compare our methods with the S2RSCS model, the state-of-the-art algorithm on this dataset. The experimental results show that CNN + FastText can outperform other approaches (CNN + GloVe, LSTM + GloVe, LSTM + FastText, S2RSCS) in term of top 1 accuracy for seven types of input data. Without using a list of keywords in the input data, CNN + GloVe/FastText can surpass other techniques. It has a bit lower performance than S2RSCS in terms of the top 3 and top 5 accuracies when using the keyword information. Finally, the combination of S2RSCS and CNN + FastText, namely S2CFT, can create a better model that bypasses all other methods by top K accuracy (K = 1,3,5,10).

In many countries worldwide, effectively estimating AQI values and levels is essential for better monitoring the air pollution around the living area. This problem has become one of the interesting research subjects for many years, and there are many applications developed for personal usages. In this work, we aim to investigate a multi-source machine learning approach to approximate the local AQI scores at users’ location in a big city. We conduct different experiments on three primary data sets: "SEPHLA-MediaEval 2019", "MNR-Air-HCM," and "MNR-HCM," collected in Ho Chi Minh City (Vietnam) and Fukuoka city (Japan). From the data sets provided, we extract different types of useful attributes for the problem: the timestamp information, the geographical data, sensor data (humidity and temperature), users’ emotion tags (such as greenness, calmness, etc.), the semantic features from images captured by users as well as the public weather data (including temperature, dew point, humidity, wind speed, and pressure) of the related cities. After that, we compare five distinct machine learning models for estimating the local AQI score and level, including Support Vector Machine [1], Random Forest [2], Extreme Gradient Boosting [3], LightGBM [4] and CatBoost [5]. We use RMSE, MAE, and R2 for measuring the performance of these approaches. The experimental results show that using random forest with sensor data, combined with public weather data, the results in AQI values regression and AQI ranks prediction can be the highest in many cases.

Partial differential equations (PDEs) have a lot of applications in different fields of research during the last decades. In this paper, we study a mesh-free deep learning method for solving PDE systems, especially for Poisson’s equations. Different from traditional techniques using finite volume or finite element method, we design suitable neural networks that can approximate solutions of a PDE hy formulating it as an optimization problem. To minimize a loss function, we use the gradient descent algorithm to obtain the neural networks’ optimal set of parameters. The experimental results show that the proposed methods can achieve promising results in solving three types of PDEs: Burgers’ equation, Laplace’s equation, and Poisson’s equation, where the mean square errors vary from 10-7 to 10-10.

Nowadays, recommendation systems play an indispensable role in many fields, including e-commerce, finance, economy, and gaming. There is emerging research on publication venue recommendation systems to support researchers when submitting their scientific work. Several publishers such as IEEE, Springer, and Elsevier have implemented their submission recommendation systems only to help researchers choose appropriate conferences or journals for submission. In this work, we present a demo framework to construct an effective recommendation system for paper submission. With the input data (the title, the abstract, and the list of possible keywords) of a given manuscript, the system recommends the list of top relevant journals or conferences to authors. By using state-of-the-art techniques in natural language understanding, we combine the features extracted with other useful handcrafted features. We utilize deep learning models to build an efficient recommendation engine for the proposed system. Finally, we present the User Interface (UI) and the architecture of our paper submission recommendation system for later usage by researchers.

The explosive rise of intelligent devices with ubiquitous connectivity have dramatically increased Internet of Things (IoT) traffic in the cloud environment and created potential attack surfaces for cyber-attacks. Traditional security approaches are insufficient and inefficient to address security threats in cloud-based IoT networks. In this vein, software defined networking (SDN), network function virtualization (NFV), and machine learning techniques introduce numerous advantages that can effectively resolve cybersecurity matters for cloud-based IoT systems. In this paper, we propose a collaborative and intelligent network-based intrusion detection system (NIDS) architecture, namely SeArch for SDN-based cloud IoT networks. It composes a hierarchical layer of intelligent IDS nodes working in collaboration to detect anomalies and formulate policy into the SDN-based IoT gateway devices to stop malicious traffic as fast as possible. We first describe a new NIDS architecture with a comprehensive analysis in terms of the system resource and path selection optimizations. Next, the system process logic is extensively investigated through main consecutive procedures, including initialization, runtime operation, and database update. Afterward, we conduct a detailed implementation of the proposed solution in an SDN-based environment and perform a variety of experiments. Finally, evaluation results of the SeArch architecture yield outstanding performance in anomaly detection and mitigation as well as bottleneck problem handling in the SDN-based cloud IoT networks in comparison with existing solutions.

With the increasing number of scientific publications as well as conferences and journals, it is often hard for researchers (especially newcomers) to find a suitable venue to present their studies. A submission recommendation system would be hugely helpful to assist authors in deciding where they can submit their work. In this paper, we propose a novel approach for a Scientific Submission Recommendation System for Computer Science (S2RSCS) by using the necessary information from the title, the abstract, and the list of keywords in given paper submission. By using tf-idf, the chi-square statistics, and the one-hot encoding technique, we consider different schemes for feature selection, which can be extracted from the title, the abstract, and keywords, to generate various groups of features. We investigate two machine-learning models, including Logistic Linear Regression (LLR) and Multi-layer Perceptrons (MLP), for constructing an appropriate recommendation engine. The experimental results show that using keywords can help to increase the performance of the recommendation model significantly. Prominently, the proposed methods outperform the previous work [1] for different groups of features in terms of top-3 accuracy. These results can give a promising contribution to the current research of the paper recommendation topic.

Food image recognition has been extensively investigated during the last decade and had multiple useful applications for monitoring food calories and analyzing people’s eating habits to ensure better health. In this paper, we study a Vietnamese food recognition system using Convolutional Neural Networks (CNNs) based features. We manually collect one dataset for Vietnamese food classification with 13 categories and 8903 images. For learning a proper food classifier, we conduct brief analytics by comparing hand-crafted features and CNNs based features (including AlexNet, GoogleNet, ResNet50, ResNet101v2, and InceptionResnetv2) and choosing top K accuracy for measuring the performance of each model. The experimental results show that InceptionResnetv2 can achieve the best performance among all these techniques. We aim at publishing our codes and datasets for giving and additional contribution to the research community related to the Vietnamese food recognition problem.

Adding appropriate punctuation marks into text is an essential step in speech-to-text where such information is usually not available. While this has been extensively studied for English, there is no large-scale dataset and comprehensive study in the punctuation prediction problem for the Vietnamese language. In this paper, we collect two massive datasets and conduct a benchmark with both traditional methods and deep neural networks. We aim to publish both our data and all implementation codes to facilitate further research, not only in Vietnamese punctuation prediction but also in other related fields. Our project, including datasets and implementation details, is publicly available at https://github.com/BinhMisfit/vietnamese-punctuation-prediction.

Portfolio optimization has been broadly investigated during the last decades and had a lot of applications in finance and economics. In this paper, we study the portfolio optimization problem in the Vietnamese stock market by using deep-learning methodologies and one dataset collected from the Ho Chi Minh City Stock Exchange (VN-HOSE) from the beginning of the year 2013 to the middle of the year 2019. We aim to construct an efficient algorithm that can find the portfolio having the highest Sharpe ratio in the next coming weeks. To overcome this challenge, we propose a novel loss function and transform the original problem into a supervised problem. The input data can be determined as a 3D tensor, while the predicted output is the unnormalized weighted proportion for each ticker in the portfolio to maximize the daily return Y of the stock market after a given number of days. We compare different deep learning models, including Residual Networks (ResNet), Long short-term memory (LSTM), Gated Recurrent Unit (GRU), Self-Attention (SA), Additive Attention (AA), and various combinations: SA + LSTM, SA + GRU, AA + LSTM, and AA + GRU. The experimental results show that the AA + GRU outperforms the rest of the methods on the Sharpe ratio and provides promising results for the portfolio optimization problem not only in Vietnam but also in other countries.

Autonomous vehicles have been investigated broadly during the last decade and predicted to decrease road fatalities by shifting control of safety-critical tasks from humans to machines. An early unsafety detection consequently becomes a key feature in every self-driving cars and trucks. In this paper, we present a promising approach for the safety prediction problem in autonomous vehicles by using one dataset collected from the competition CMDC 2019, which can capture multiple safe or unsafe situations from a front car camera put in different autonomous buses. We consider various ways to extract potential features from images provided and apply numerous machine learning techniques to learn an efficient detection algorithm. The experimental results show that by combining Histogram-of-Gradients (HOG) features as well as deep-learning ones computed from both ResNet50 and our proposed deep neural networks (MRNets), we can achieve an auspicious performance in terms of both micro-averaged F1-score and macro-averaged F1-score. The outcome of our papers can give an additional contribution to the current study of the problem.

The influences of pH value on Pb(II) removal by the biopolymer extracted from activated sludge were investigated from various perspectives, including removal rate, functional groups, binding site number between the biopolymer and Pb(II), and removal distribution. With the system pH value rising from 4.0 to 9.0, the total removal rate of Pb(II) without and with adding the biopolymer increased from 1.94% to 86.5% and from 32.2% to 95.4%, respectively. From the analyses of Fourier transform infrared (FTIR) spectroscopy and two-dimensional correlation spectroscopy (2D-COS), the continuous dissociation and deprotonation of functional groups in the biopolymer with rising system pH value promoted the biosorption removal of Pb(II) via complexation and ion exchange. According to the three-dimensional excitation emission matrix (3D-EEM) fluorescence spectrum, two protein-like fluorescence peaks of A (Ex/Em = 280 nm/326–338 nm) and B (Ex/Em = 220–230 nm/324–338 nm) were identified in the biopolymer. The binding site numbers, which were obtained via fluorescence quenching titration experiments, increased first and then decreased with rising system pH value for both peaks A and B. Biosorption removal of Pb(II) was dominant under acid conditions, but biosorption and Pb(OH) 2 precipitation co-existed under neutral and alkaline conditions. The biosorption and the Pb(OH) 2 precipitation were interestingly found to interfere with each other under alkaline conditions. The decreases in both binding site number and biosorption removal rate indicated that Pb(OH) 2 precipitation excelled the biosorption in the competition of capturing Pb(II).

Two new series of metal–organic frameworks (MOFs), termed M-VNU-74-I and -II (where M = Mg, Ni, Co; VNU = Vietnam National University) were designed to expand the methanol uptake capacities with polar amide functionalities. The resulting MOFs, isoreticular to MOF-74, exhibited high porosity (up to 3000 m2 g–1) as well as the highest reported methanol uptake [>1 g g–1 or >400 cm3 cm–3]. As a representative example, Mg-VNU-74-II was shown to maintain a remarkably high stability and methanol mass transfer capacity for at least 42 ad/desorption cycles (3 days). Indeed, these findings highlight the potential of such materials for practical use in adsorption heat pump applications.

Building an intelligent system for automatically composing music like human beings has been actively investigated during the last decade. In this work, we propose a new approach for automatically creating hymns by training a variational attention model from a large collection of religious songs. We compare our method with two other techniques by using Seq2Seq and attention models and measure the corresponding performance by BLEU-N scores, the entropy, and the edit distance. Experimental results show that the proposed method can achieve a promising performance that is able to give an additional contribution to the current study of music formulation. Finally, we publish our dataset online for further research related to the problem.

This paper provides a fog-based approach to solving the traffic light optimization problem which utilizes the Adaptive Traffic Signal Control (ATSC) model. ATSC systems demand the ability to strictly reflect real-time traffic state. The proposed fog computing framework, namely FogFly, aligns with this requirement by its natures in location-awareness, low latency and affordability to the changes in traffic conditions. As traffic data is updated timely and processed at fog nodes deployed close to data sources (i.e., vehicles at intersections) traffic light cycles can be optimized efficiently while virtualized resources available at network edges are efficiently utilized. Evaluation results show that services running in FogFly produce better performance comparing to those in cloud computing approaches.

We study a dynamic spectrum access situation where, in each time slot, a single cognitive agent decides to either stay idle or access one of the N frequency channels based on its sensing of the whole spectrum. The channels are occupied or vacant according to N independent nonidentical 2-state Markov chains. We prove that the optimal access policy can easily be found if the state transition probabilities of all channels are known. When the agent has no knowledge about the channel model, we propose to use the deep Q-learning method to learn a state-action value function that determines an access policy from the observed states of all channels. In this method, the optimal Q-function is approximated with a neural network of all dense layers that is trained via experience replay. We demonstrate through experiments that the learning-based policies consistently achieve performances that are close to the optimal ones.

In this paper, we propose a deep learning based system for food recognition from personal life archive im- ages. The system first identifies the eating moments based on multi-modal information, then tries to focus and enhance the food images available in these moments, and finally, exploits GoogleNet as the core of the learning process to recognise the food category of the images. Preliminary results, experimenting on the food recognition module of the proposed system, show that the proposed system achieves 95.97% classification accuracy on the food images taken from the personal life archive from several lifeloggers, which potentially can be extended and applied in broader scenarios and for different types of food categories.

In this paper, we build an emotion prediction system by only using heart rate signals. This system can easily support any heart-rate ‘sensor and smartphone to enhance users’ experience. We collect heart-rate data from registered users from a heart-rate sensor by building an Android application, namely Emotion and Heart Rate Collection. We analyze different kinds of feature vectors and compare various supervised learning models, including k-Nearest Neighbor (kNN), Support Vector Machine (SVM), and decision tree. The experiments show that using SVM can achieve the highest performance in comparison with other approaches.

In this paper, we investigate Vietnamese food recognition problem and compare different methods by using traditional hand-crafted features and convolutional neural networks (CNNs). To this aim, we first introduce one dataset including five popular Vietnamese dishes with more than 2000 images. We then conduct various experiments on that dataset and analyze the performance of each approach. The experiments show that CNNs outperform the hand-crafted techniques by a significant margin.

Phosphorus (P) is an essential nutrient that affects the growth and metabolism of microalgal biomass. Despite the obvious importance of P, the dynamics of how it is taken up and distributed in microalgae are largely undefined. In this study, we tracked the fate of P during batch growth of the cyanobacterium Synechocystis sp. PCC 6803. We determined the distribution of P in intracellular polymeric substances (IPS), extracellular polymeric substances (EPS), and soluble microbial products (SMP) for three initial ortho-phosphate concentrations. Results show that the initial P concentration had no impact on the production of biomass, SMP, and EPS. While the initial P concentration affected the rate and the timing of how P was transformed among internal and external forms of inorganic P (IP) and organic P (OP), the trends were the same no matter the starting P concentration. Initially, IP in the bulk solution was rapidly and simultaneously adsorbed by EPS (IPEPS) and taken up as internal IP (IPint). As the bulk-solution's IP was depleted, desorption of IPEPS became the predominant source for IP that was taken up by the growing cells and converted into OPint. At the end of the 9-d batch experiments, almost all P was OP, and most of the OP was intracellular. Based on all of the results, we propose a set of transformation pathways for P during the growth of Synechocystis. Key is that EPS and intracellular P pool play important and distinct roles in the uptake and storage of P.

Emotion recognition has many useful applications in daily lives. In this paper, we present a potential approach to detect human emotion in real time. For any face detected in camera, we extract the corresponding facial landmarks and examine different kinds of features and models for predicting human emotion. The experiments show that our proposed system can naturally detect human emotion in real time and achieve an average accuracy about 70.65%.

In this paper, we investigate the problem of user movement prediction from historical location data. We create an Android application, namely Movement Predictor, that can help to collect location data from registered users by Global Positioning System (GPS) signals. We analyze different kinds of feature vectors and compare three supervised learning models: Markov model, Support Vector Machine (SVM), and decision tree. The experiments show that SVM model can achieve the highest performance (with an accuracy 92%) in comparison with other approaches.

Set inversion is a classical problem in control theory that has many important applications in various fields of science and engineering. The state-of-the-art method for solving this problem, Set Inverter Via Interval Analysis (SIVIA), usually does not work well in high dimensions and often fails to recover sets with complicated structures. In this work, we propose a new approach to the problem of set inversion, which employs techniques from machine learning to resolve these issues. Our algorithm can handle problems in high dimensions and achieve the same level of accuracy with fewer data points compared to SIVIA. We illustrate the performance of our method in various simulation studies and apply it to investigate the dynamics of the 17th-century plague in Eyam village, England.

The quantification and effects of system pH value on the interactions between Pb(II) and the biopolymer in activated sludge were investigated. The biopolymer had two protein-like fluorescence peaks (Ex/Em = 280 nm/326–338 nm for peak A; Ex/Em = 220–230 nm/324–338 nm for peak B). The fluorescence intensities of peak B were higher than those of peak A. The fluorophores of both peaks could be largely quenched by Pb(II), and the quencher dose for peak B was about half of that for peak A. The modified Stern-Volmer equation well depicted the fluorescence quenching titration. The quenching constant (Ka) values for both peaks decreased with rising system pH value, and then sharply decreased under alkaline conditions. It could be attributed to that the alkaline conditions caused the reduction of available Pb(II) due to the occurrence of Pb(OH)2 sediments. The Ka values of peak B were bigger than those for peak A at the same system pH values. Accordingly, the aromatic proteins (peak B) played a key role in the interactions between metal ions and the biopolymer.