Search

Although many variable decomposition methods for cooperative co-evolution (CC) have been proposed, researches on scalable and efficient decomposition have rarely been done, particularly, for the large-scale global optimization (LSGO) problems. In this paper, we propose an efficient variable interdependency identification and decomposition method, called EVIID. Different from existing studies focusing on only limited scale efficient or accurate variable decomposition, our purpose is to develop a scalable variable decomposition method with high efficiency and accuracy even on very high-dimensional problems. EVIID utilizes three core strategies: a binary variable space search, a dynamic perturbation caching, and a pre-variable sorting. Their synergy effect enables scalable and efficient variable decomposition without sacrificing decomposition accuracy by pruning many redundant computations required to identify interdependencies among decision variables. In comprehensive experiments, EVIID showed highly scalable decomposition ability on 1000 to 10,000 dimensional benchmark problems compared against the state-of-the-art variable decomposition methods. Moreover, when EVIID was embedded into practical CC frameworks, it showed good optimization performance and also fast convergence.

This work uses sequence-to-sequence (seq2seq) models pre-trained on monolingual corpora for machine translation. We pre-train two seq2seq models with monolingual corpora for the source and target languages, then combine the encoder of the source language model and the decoder of the target language model, i.e., the cross-connection. We add an intermediate layer between the pre-trained encoder and the decoder to help the mapping of each other since the modules are pre-trained completely independently. These monolingual pre-trained models can work as a multilingual pre-trained model because one model can be cross-connected with another model pre-trained on any other language, while their capacity is not affected by the number of languages. We will demonstrate that our method improves the translation performance significantly over the random baseline. Moreover, we will analyze the appropriate choice of the intermediate layer, the importance of each part of a pre-trained model, and the performance change along with the size of the bitext.

As the performance of devices that conduct large-scale computations has been rapidly improved, various deep learning models have been successfully utilized in various applications. Particularly, convolution neural networks (CNN) have shown remarkable performance in image processing tasks such as image classification and segmentation. Accordingly, more stable and robust optimization methods are required to effectively train them. However, the traditional optimizers used in deep learning still have unsatisfactory training performance for the models with many layers and weights. Accordingly, in this paper, we propose a new Adam-based hybrid optimization method called HyAdamC for training CNNs effectively. HyAdamC uses three new velocity control functions to adjust its search strength carefully in term of initial, short, and long-term velocities. Moreover, HyAdamC utilizes an adaptive coefficient computation method to prevent that a search direction determined by the first momentum is distorted by any outlier gradients. Then, these are combined into one hybrid method. In our experiments, HyAdamC showed not only notable test accuracies but also significantly stable and robust optimization abilities when training various CNN models. Furthermore, we also found that HyAdamC could be applied into not only image classification and image segmentation tasks.

Gaze following is the task of detecting the point of attention of where a third person gaze is staring in a single image. Existing studies have made some modifications to architectures or have additionally learned the gaze angle, and have achieved notable performances. However, when a complex scene is given, the methods generally predict incorrect locations because of the lack of depth information in an RGB image. In this paper, we propose a novel three-stage deep neural networks algorithm to tackle such challenging scenes using a depth map. We achieve state-of-the-art performance on the GazeFollow dataset and examine possibilities for the research of depth information in image interpretation. Moreover, a qualitative comparison shows that our method works stably and accurately for complex scenes similar to those found in real-world photographs.

Skeleton-based action recognition using a Graph Convolutional Network (GCN) achieved remarkable results by reconstructing a person's skeleton into a graph. However, there are fundamental problems with existing GCN-based models. Generally, human action is greatly affected by the tools used, but in traditional GCN models, action recognition is performed without using the information in the tools. For example, a person holding a pen is limited to the act of writing. A person holding a ball is limited to the action of throwing or receiving the ball. In other words, it is an inaccurate method to judge action only by recognizing the movements of bones used in existing methods. Therefore, a graph was made to reflect the information on the tool. We identify tool-information using the LSTM classifier and propose GCNs for skeleton-based action recognition with LSTM using tool-information. Additionally, we apply a new graph construction and utilize the Learnable Adjacency Matrix. The proposed method is applied to the existing model and comparative evaluation was performed between the model with and without the applied algorithm. The evaluations showed consistent performance improvement, and the proposed method applied to the baseline models achieved state-of-the-art performance.

In this paper, we propose several methods to improve the performance of multiple object tracking (MOT), especially for humans, in dynamic environments such as robots and autonomous vehicles. The first method is to restore and re-detect unreliable results to improve the detection. The second is to restore noisy regions in the image before the tracking association to improve the identification. To implement the image restoration function used in these two methods, an image inference model based on SRGAN (super-resolution generative adversarial networks) is used. Finally, the third method includes an association method using face features to reduce failures in the tracking association. Three distance measurements are designed so that this method can be applied to various environments. In order to validate the effectiveness of our proposed methods, we select two baseline trackers for comparative experiments and construct a robotic environment that interacts with real people and provides services. Experimental results demonstrate that the proposed methods efficiently overcome dynamic situations and show favorable performance in general situations.

With recent advances in biotechnology and sequencing technology, the microbial community has been intensively studied and discovered to be associated with many chronic as well as acute diseases. Even though a tremendous number of studies describing the association between microbes and diseases have been published, text mining methods that focus on such associations have been rarely studied. We propose a framework that combines machine learning and natural language processing methods to analyze the association between microbes and diseases. A hierarchical long short-term memory network was used to detect sentences that describe the association. For the sentences determined, two different parse tree-based search methods were combined to find the relation-describing word. The ensemble model of constituency parsing for structural pattern matching and dependency-based relation extraction improved the prediction accuracy. By combining deep learning and parse tree-based extractions, our proposed framework could extract the microbe-disease association with higher accuracy. The evaluation results showed that our system achieved an F-score of 0.8764 and 0.8524 in binary decisions and extracting relation words, respectively. As a case study, we performed a large-scale analysis of the association between microbes and diseases. Additionally, a set of common microbes shared by multiple diseases were also identified in this study. This study could provide valuable information for the major microbes that were studied for a specific disease. The code and data are available at https://github.com/DMnBI/mdi_predictor.

Video captioning is known as a method to summarize or explain a video. However, it is very difficult to create sports commentaries, which are running scene-by-scene descriptions, for sports videos by using conventional video captioning. This is the case because sports commentary requires not only specific and varied information about every scene, such as player action descriptions in baseball, but also background knowledge and dynamic at-bats statistics that are not found in the video. We propose a new system to automatically generate commentary for baseball games. In our system, given real-time baseball videos, suitable descriptions are relayed using four deep-learning models (i.e., a scene classifier, player detector, motion recognizer, and pitching result recognizer) integrated with domain ontology. Using these four deep-learning models, pieces of information about "who is doing what in which area of the field" and "what results are expected" are obtained. This approach is used to select an appropriate template, which is combined with baseball ontology knowledge for the generation of commentary. We evaluate our system using baseball games from the KBO (Korea Baseball Organization) League's 2018 season. As a result of the evaluation, we identify that our system1 can act as a commentator for baseball games.

Recently, several studies have focused on image-to-image translation. However, the quality of the translation results is lacking in certain respects. We propose a new image-to-image translation method to minimize such shortcomings using an auto-encoder and an auto-decoder. This method includes pre-training two auto-encoders and decoder pairs for each source and target image domain, cross-connecting two pairs and adding a feature mapping layer. Our method is quite simple and straightforward to adopt but very effective in practice, and we experimentally demonstrated that our method can significantly enhance the quality of image-to-image translation. We used the well-known cityscapes, horse2zebra, cat2dog, maps, summer2winter, and night2day datasets. Our method shows qualitative and quantitative improvements over existing models.

Zero-shot recognition (ZSR) aims to perform visual classification by category in the absence of training samples. The focus in most traditional ZSR models is using semantic knowledge about familiar categories to represent unfamiliar categories with only the visual appearance of an unseen object. In this research, we consider not only visual information but context to enhance the classifier&rsquo, s cognitive ability in a multi-object scene. We propose a novel method, contextual inference, that uses external resources such as knowledge graphs and semantic embedding spaces to obtain similarity measures between an unseen object and its surrounding objects. Using the intuition that close contexts involve more related associations than distant ones, distance weighting is applied to each piece of surrounding information with a newly defined distance calculation formula. We integrated contextual inference into traditional ZSR models to calibrate their visual predictions, and performed extensive experiments on two different datasets for comparative evaluations. The experimental results demonstrate the effectiveness of our method through significant enhancements in performance.

Many researchers recently paid attention to applying deep learning to collaborative recommendation. Especially, RNN(Recurrent Neural Network)-based recommender system was shown to learn users' interest and preference from temporal sequences of users' movie consumption records, and they could make better recommendation compared to conventional collaborative recommendation. In this work, we present an embedding-weight tying approach to RNN-based recommendation in order to improve the performance of movie recommender system more. In experiments, our approach outperforms existing RNN-based recommendation as well as currently popular collaborative recommendation in terms of short-term prediction success(sps) and recall.

We developed a game based content for learning to minimize boolean functions using Karnaugh Map. With interesting game based scenarios and an intelligent tutoring module, our content could give adaptive feedbacks to learners as needed in order to make the learners well-motivated and immersed in learning context. We performed an experiment for three student groups (one experimental group and two control groups) which were learning the minimization of boolean functions at a vocational high school in Korea, and compared our content aided learning to other conventional ones for evaluating the learning effectiveness. In this experiment, we identified that our content aided learning was more effective than other ones significantly in terms of both academic achievement and learning attitude. More specifically, our content was very useful for learners to improve academic attitude and learning habit.

Classifying semantic relations between entity pairs in sentences is an important task in natural language processing (NLP). Most previous models applied to relation classification rely on high-level lexical and syntactic features obtained by NLP tools such as WordNet, the dependency parser, part-of-speech (POS) tagger, and named entity recognizers (NER). In addition, state-of-the-art neural models based on attention mechanisms do not fully utilize information related to the entity, which may be the most crucial feature for relation classification. To address these issues, we propose a novel end-to-end recurrent neural model that incorporates an entity-aware attention mechanism with a latent entity typing (LET) method. Our model not only effectively utilizes entities and their latent types as features, but also builds word representations by applying self-attention based on symmetrical similarity of a sentence itself. Moreover, the model is interpretable by visualizing applied attention mechanisms. Experimental results obtained with the SemEval-2010 Task 8 dataset, which is one of the most popular relation classification tasks, demonstrate that our model outperforms existing state-of-the-art models without any high-level features.

Recently, due to the widespread on-line availability of syntactically annotated text corpora, some automated tools for searching in such text corpora have gained great attention. Generally, those conventional corpus search tools use a decomposition-matching-merging method based on relational predicates for matching a tree pattern query to the desired parts of text corpora. Thus, their query formulation and expressivity are often complicated due to poorly understood query formalisms, and their searching tasks may require a big computational overhead due to a large number of repeated trials of matching tree patterns. To overcome these difficulties, we present TPEMatcher, a tool for searching in parsed text corpora. TPEMatcher provides not only an efficient way of query formulation and searching but also a good query expressivity based on concise syntax and semantics of tree pattern query. We also demonstrate that TPEMatcher can be effectively used for a text mining in practice with its useful interface providing in-depth details of search results.

In this paper, we propose an adaptive and collaborative recommendation method based on content type, which can enhance performance considerably in practice. Conventional collaborative recommendations are troubled with no or little effective rating information for newly comers or even some old users so that they often work poorly. In order to relax such cold start or sparse rating information problems, we employ a user-content type matrix with relatively higher density than commonly-used user-content matrix. By using user-content_type matrix, we evaluate user's preference for a content type and then reflect it to the final prediction of content preference in collaborative recommendation. In such a way, our method adaptively combines content preference with content type preference. In experiments, we identify notable performance improvement compared to traditional collaborative recommendation methods in terms of MAE (Mean Absolute Error) and coverage.

Question and answer archives have become useful information resources with increase of community based question and answer service. For effective question and answering system, it is important to find semantically similar questions and retrieve its answer from the archive for user's question. In this paper, we propose a weighted combination of retrieval models for question and answer archives. In contrast to well-known translation based language model, the proposed model reflects significance of each word in user's question by giving variant weight to the word depending on its part-of-speech. In experiment, our model improves performance considerably when compared to conventional ones.

In order to develop effective question and answering (Q&A) system, it is necessary to analyze syntactic structures of sentence and extract a set of semantically important words (tuple) from them. In this paper, we propose Tree Pattern Expression (TPE) based tuple extraction method for high-performance Q&A system. To extract tuples from target sentences more accurately than conventional Q&A systems, TPE and its fast pattern matching algorithm are used. A tuple consists of subject, predicate, object, preposition and its object, which may contain a correct answer part for a given question. In experiments, our TPE based method performs much better than traditional ones for Q&A system, as we expected, because our method elicits key structures of complicated sentences elaborately and thus extracts their associated tuples more accurately.

Since documents on the Web are naturally partitioned into many text databases, the efficient document retrieval process requires identifying the text databases that are most likely to provide relevant documents to the query and then searching for the identified text databases. In this paper, we propose a neural net based approach to such an efficient document retrieval. First, we present a neural net agent that learns about underlying text databases from the user's relevance feedback. For a given query, the neural net agent, which is sufficiently trained on the basis of the BPN learning mechanism, discovers the text databases associated with the relevant documents and retrieves those documents effectively. In order to scale our approach with the large number of text databases, we also propose the hierarchical organization of neural net agents which reduces the total training cost at the acceptable level. Finally, we evaluate the performance of our approach by comparing it to those of the conventional well-known approaches.

As large numbers of text databases have become available on the Web, many efforts have been made to solve the text database discovery problem: finding which text databases (out of many candidates) are most likely to provide relevant documents to a given query. In this paper, we propose a neural net based approach to this problem. First, we present a neural net agent that learns about underlying text databases from the user's relevance feedback. For a given query, the neural net agent, which is sufficiently trained on the basis of the backpropagation learning mechanism, discovers the text databases associated with the relevant documents and retrieves those documents effectively. In order to scale our approach with the large number of text databases, we also propose the hierarchical organization of neural net agents which reduces the total training cost at the acceptable level. Finally, we evaluate the performance of our approach by comparing it to those of the conventional well-known statistical approaches.

Introduction Registering a 3-dimensional (3D) facial surface scan to a cone-beam computed tomography (CBCT) scan has various advantages. One major advantage is to compensate for the inaccuracy of the CBCT surface data. However, when registering CBCT and 3D facial scans, changes in facial expression, spatial soft-tissue changes, and differences in the patient's positioning can decrease the accuracy of the registration. In this study, we introduce a new 3D facial scanner that is combined with a CBCT apparatus. Our goal was to evaluate the registration accuracy of CBCT and 3D facial scans, which were taken with the shortest possible time between them. Methods The experiment was performed with 4 subjects. Each patient was instructed to hold as still as possible while the CBCT scan was taken, followed immediately by the 3D facial surface scan. The images were automatically registered with software. The accuracy was measured by determining the degree of agreement between the soft-tissue surfaces of the CBCT and the 3D facial images. Results The average surface discrepancy between the CBCT facial surface and 3D facial surface was 0.60 mm (SD, 0.12 mm). Registration accuracy was also visually verified by toggling between the images of the CBCT and 3D facial surface scans while rotating the registered images. Conclusions Registration of consecutively taken CBCT and 3D facial images resulted in reliable accuracy.

In this paper, we propose a novel CAT (Computerized Adaptive Testing) system based on Bayesian network. Our novel system makes good use of topology and probabilistic inference algorithm of Bayesian network to efficiently estimate proficiency of learner and also give an adaptive learning guide when needed. From several experiments, we identified that our system could considerably improve proficiency-estimation performance when compared with conventional CAT methods.

For accurate and consistent fault diagnosis for those systems where the amount of well-defined diagnosis knowledge is vast and the sequence of steps required to identify faults is very long, an approach to building an expert system may be suggested. In this paper we introduce an expert system for diagnosing faults in the Vulcan system of air defense artilleries. First we briefly describe an expert system shell, HEXPERT, used for developing the system and then explain the process to formulate the knowledge base which includes the knowledge acquisition and the knowledge analysis steps. Finally the system developed is illustrated for three cases associated with three different faults in the Vulcan.

In this paper, we propose a game based learning content supporting intelligent tutoring functionality. The learning objective of our game based learning content is to understand the Karnaugh Map which is generally utilized to simplify boolean functions. Our game based learning content well-motivates learners with interesting game based scenarios and also its intelligent tutoring module gives learners some adaptive feedbacks such as hints and explanations as needed, while maintaining learners immersed in learning context. In the experiment of our game based learning content, we identified significant improvement in terms of learning effectiveness by analyzing the test results of two (experimental and control) student groups learning the simplification of boolean functions.

The objective of this study was to mark the points of emphasis in hidden dental caries from 15 subjects’ dental X-ray images. Each image is an 8-bit grayscale image taken by digital image radiography. For each caries lesion diagnosed, a mean pixel value was obtained in digital subtraction radiographs using a scale ranging from 0 to 255 gray values. Hence, the purpose of this study was to design a novel detection method of proximal caries, based on pixel gray values in digital subtraction radiography images from a patient’s single dental image, used for monitoring of dental caries. As a result, in digital subtraction radiography, the image is used to mark the points of emphasis in hidden dental caries using a novel detection method of proximal caries from a patient’s single dental image.