Vol. 48, No. 7,
Jul. 2021
Digital Library
Xpass: NUMA-aware Persistent Memory Disaggregation
Jaeyoun Nam, Hokeun Cha, ByeongKeon Lee, Beomseok Nam
http://doi.org/10.5626/JOK.2021.48.7.735
The disaggregation method is used for efficient resource management in large-scale data centers, where each server consists of NUMA nodes. In the NUMA architecture, the latency difference between the remote and local access is known to be significant. In particular, remote NUMA access to persistent memory is even higher than DRAM. In this study, we propose Xpass, a memory disaggregation framework that considers the locality of NUMA architecture in a persistent memory disaggregation system. Xpass uses the dynamic hash table - CCEH to manage cached pages, and proposes a segment split algorithm that considers load balancing between the NUMA nodes in a NUMA environment.
Opt Tree: Write Optimized Tree Using Optane DCPM Internal Buffer
http://doi.org/10.5626/JOK.2021.48.7.742
Intel’s Optane DC Persistent Memory, a recently commercialized non-volatile byte-addressable memory, has an internal buffer of 256 bytes called XPLine, which processes memory access commands in units of cache lines or words. In this paper, we propose Opt Tree, a novel byte-addressable persistent index that utilizes the internal buffer of the Optane DCPM. Opt Tree divides the tree node into several small blocks of 256 bytes. For insertions and searches, Opt Tree accesses only two blocks. In our performance study, Opt Tree shows better insertion performance than the existing persistent indexes through its internal buffer-friendly design.
Hybrid Tracing Based on Primitives and Voxels for Real-Time Global Illumination
http://doi.org/10.5626/JOK.2021.48.7.748
The voxel cone tracing algorithm based on the sparse octree stores indirect lighting attributes in a volume, and efficiently traces the volume for indirect illumination and reflections. However, as the volume resolution becomes higher, the amounts of memory and performance overhead significantly increase for the voxel cone tracing. In this paper, we propose a hybrid octree data structure, where sparse leaf nodes use primitive lists while dense leaf nodes still use voxels, and a hybrid tracing algorithm that adaptively performs the cone tracing and ray tracing. Our real-time solution outperforms the existing voxel cone tracing algorithm for high-quality indirect illumination.
GPU-Based High-Precision Adaptive Vertex Depth Rendering
http://doi.org/10.5626/JOK.2021.48.7.756
The Z-buffer algorithm is a standard technique for visibility determination in the modern 3D graphics pipeline. However, its limited precision can lead to z-fighting. This can be mitigated by adjusting per-object projection matrices, but its per-object iteration and draw calls become costly for complex scenes. In this paper, we introduce a GPU-based high-precision adaptive vertex depth rendering technique. The technique adjusts the z coordinates of the vertices in the clip space in the vertex shader. The adjusted vertices are adaptively biased depending upon the location of the objects through the rendering pipeline. Our technique can be applied to problems relating to depth precision including GPU occlusion culling and shadow mapping.
Prediction of Fine Dust in Gyeonggi-do Industrial Complex using Machine Learning Methods
Dong-Jun Won, Sun-Kyum Kim, Yeonghun Kim, Gyuwon Song
http://doi.org/10.5626/JOK.2021.48.7.764
Recently, research on fine dust has been conducted through various prediction techniques. However, currently the research focused on PM10 concentration prediction, and thus it is necessary to develop a model capable of predicting PM2.5 concentration. In this paper, we have collected air quality, weather, and traffic of the Banwol Shihwa National Industrial Complex in the recent two years. The significance of the variable been identified through correlation analysis and regression analysis among PM2.5 and PM10, SO₂, NO₂, CO, O₃, temperature, humidity, wind direction, wind speed, precipitation, road section vehicle speed for each vehicle. Next, the data has been used to predict PM2.5 concentration based on time in the industrial complex. Through the artificial intelligence techniques, Random Forest, XGBoost, LightGBM, Deep neural network and Voting models, PM2.5 concentration industrial complexes been predicted on an hourly basis, and comparative analysis been conducted based on RMSE. As a result of prediction, RMSE was 6.27, 6.41, 6.22, 6.64, and 6.12, respectively, and each technique showed very high performance compared to 10.77 of the technique predicted by Air Korea.
Alpha-Integration Pooling for Convolutional Neural Networks
http://doi.org/10.5626/JOK.2021.48.7.774
Convolutional neural networks (CNNs) have achieved remarkable performance in many applications, especially in image recognition tasks. As a crucial component of CNNs, sub-sampling plays an important role for efficient training or invariance property, and max-pooling and arithmetic average-pooling are commonly used sub-sampling methods. In addition to the two pooling methods, however, there are many other pooling types, such as geometric average, harmonic average, among others. Since it is not easy for algorithms to find the best pooling method, usually the pooling types are predefined, which might not be optimal for different tasks. As other parameters in deep learning, however, the type of pooling can be driven by data for a given task. In this paper, we propose α-integration pooling (αI-pooling), which has a trainable parameter α to find the type of pooling. αI-pooling is a general pooling method including max-pooling and arithmetic average-pooling as a special case, depending on the parameter α. Experiments show that αI-pooling outperforms other pooling methods, in image recognition tasks. Also, it turns out that each layer has a different optimal pooling type.
A Dynamic Gesture Recognition System based on Trajectory Data of the Motion-sphere
Jaeyeong Ryu, Adithya B, Ashok Kumar Patil, Youngho Chai
http://doi.org/10.5626/JOK.2021.48.7.781
Recently, dynamic gesture recognition technology, which belongs to human-computer interaction (HCI), has received much attention. This is because the interface configuration for utilizing the system is simple and it is possible to communicate quickly. In this paper, we used a new input data format for the dynamic gesture recognition system and conducted research to improve the recognition accuracy. In the existing dynamic gesture recognition system, the position data and the rotation data of the joint are mainly used. In the proposed system, motion-sphere trajectory data are used. Motion-sphere expresses motion intuitively as a technique for visualizing movement. In the motion-sphere, the expression is composed of the trajectory and twist angle. In this paper, the trajectory of the motion-sphere is used as input data of the dynamic gesture recognition system. The validity of the trajectory data used is verified through the dynamic gesture recognition accuracy comparison. In the experiment, we experimented on two cases. The first cases were conducted by using measured quaternion data. The other experiments used open motion data. Both experiments conducted cognitive accuracy tests, and each experiment yielded high cognitive accuracy.
Dimensional Sentiment Analysis of Korean Text using Data Balancing
http://doi.org/10.5626/JOK.2021.48.7.790
Compared with most studies on categorical sentiment analysis which aims to represent emotional states as a small set of emotion categories, there have been fewer studies on dimensional sentiment analysis which treats sentiment analysis as a regression problem because of the shortage of data. Recently, the National Information Society Agency (NIA) released open data, Multimodal Video Data, through their web site, AI Hub. Using this data, we experimented with dimensional sentiment analysis of Korean text. For this purpose, we used CNN which is one of the conventional deep learning models in NLP. We also verified that data balancing could improve the performance of models. The results show that the model trained on Multimodal Video Data performs well enough to show that the data should be useful for dimensional sentiment analysis of Korean text and that with data balancing the model can perform better in spite of their fewer training data.
RNN model for Emotion Recognition in Dialogue by incorporating the Attention on the Other’s State
http://doi.org/10.5626/JOK.2021.48.7.802
Emotion recognition has increasingly received much attention in artificial intelligence, lately. In this paper, we present an RNN model that analyzes and identifies a speaker’s emotion appeared through utterances in conversation. There are two kinds of speaker considered context, self-dependency and inter-speaker dependency. In particular, we focus more on inter-speaker dependency by considering that the state context information of the relative speaker can affect the emotions of the current speaker. We propose a DialogueRNN based model that adds a new GRU Cell for catching inter-speaker dependency. Our model shows higher performance than the performances of DialogueRNN and its three variants on multiple emotion classification datasets.
Semantic Face Transformations for Attacking Deep Neural Networks and Improving Robustness
http://doi.org/10.5626/JOK.2021.48.7.809
Deep neural networks(DNNs) have achieved great successes in various vision fields such as autonomous driving, face recognition, and object detection. However, a well-trained network can be manipulated if the input of the deep neural networks is disturbed by perturbations. Currently a common attack method is by adding perturbations to the pixel space of images by limiting the Lp-norm of the perturbations. Pixel-based transformations are easily detected by the naked eye so a realistic effective attack can be a method of disturbing the network by unnaturally transforming the image. In this paper, we proposed a new attack method to use natural color transformation through the segmentation of face images. We generated face transformation images based on semantic face transformation and conducted comprehensive experiments to show that using our face transformation reduced the accuracy rate of the classification network. Our face transformation images were also used for robustness training of the neural network. The robustness of the deep neural network was improved.
Combining Sentiment-Combined Model with Pre-Trained BERT Models for Sentiment Analysis
http://doi.org/10.5626/JOK.2021.48.7.815
It is known that BERT can capture various linguistic knowledge from raw text via language modeling without using any additional hand-crafted features. However, some studies have shown that BERT-based models with an additional use of specific language knowledge have higher performance for natural language processing problems associated with that knowledge. Based on such finding, we trained a sentiment-combined model by adding sentiment features to the BERT structure. We constructed sentiment feature embeddings using sentiment polarity and intensity values annotated in a Korean sentiment lexicon and proposed two methods (external fusing and knowledge distillation) to combine sentiment-combined model with a general-purpose BERT pre-trained model. The external fusing method resulted in higher performances in Korean sentiment analysis tasks with movie reviews and hate speech datasets than baselines from other pre-trained models not fused with sentiment-combined models. We also observed that adding sentiment features to the BERT structure improved the model’s language modeling and sentiment analysis performance. Furthermore, when implementing sentiment-combined models, training time and cost could be decreased by using a small-scale BERT model with a small number of layers, dimensions, and steps.
EFA-DTI: Prediction of Drug-Target Interactions Using Edge Feature Attention
Erkhembayar Jadamba, Sooheon Kim, Hyeonsu Lee, Hwajong Kim
http://doi.org/10.5626/JOK.2021.48.7.825
Drug discovery is a high-level field of research requiring the coordination of disciplines ranging from medicinal chemistry, systems biology, structural biology, and increasingly, artificial intelligence. In particular, drug-target interaction (DTI) prediction is central to the process of screening for and optimizing candidate substances to treat disease from a nearly infinite set of compounds. Recently, as computer performance has developed dramatically, studies using artificial intelligence neural networks have been actively conducted to reduce the cost and increase the efficiency of DTI prediction. This paper proposes a model that predicts an interaction value between a given molecule and protein using a learned molecule representation via Edge Feature Attention-applied Graph Net Embedding with Fixed Fingerprints and a protein representation using pre-trained protein embeddings. The paper describes architectures, experimental methods, and findings. The model demonstrated higher performance than DeepDTA and GraphDTA, which had previously demonstrated the best performance in DTI studies.
An Analysis of the Performance Interference among Column Families in RocksDB
Hoyoung Lee, Minho Lee, Young Ik Eom
http://doi.org/10.5626/JOK.2021.48.7.835
RocksDB is one of the representative LSM-tree-based key-value stores and provides a column family feature to classify key-value pairs based on the characteristics of data. Each column family has its write buffer and manages the classified key-value pairs with it, whereas column families have to share a single WAL file for data consistency. However, sharing the WAL file induces performance interference among the column families and reduces the write performance of RocksDB. In this paper, we have analyzed the performance degradation of RocksDB caused by performance interference among multiple column families. Consequently, we measured the write performance of RocksDB by varying the size of WAL file and the number of column families. Experimental results clearly show that the write performance of RocksDB decreases by up to 57.08% according to the size of the WAL file and the number of the constructed column families.
A Method for Cancer Prognosis Prediction Using Gene Embedding
http://doi.org/10.5626/JOK.2021.48.7.842
Identifying prognostic genes and using them to predict the prognosis of cancer patients can help provide them with more effective treatments. Many methods have been proposed to identify prognostic genes and predict cancer prognosis, and recent studies have focused on machine learning methods including deep learning. However, applying gene expression data to machine learning methods has the limitations of a small number of samples and a large number of genes. In this study, we additionally use a gene network to generate many random gene paths, which we used for training the model, thereby compensating for the small sample problem. We identified the prognostic genes and predicted the prognosis of patients using the gene expression data and gene networks for five cancer types and confirmed that the proposed method showed better predictive accuracy compared to other existing methods, and good performance on small sample data.
HTTP/3 Stream Prioritization based on Web Object Dependency
http://doi.org/10.5626/JOK.2021.48.7.850
HTTP/3 is an application layer protocol that includes new features to meet the needs of the modern web. IETF standardization of HTTP/3 has come to its final stage. HTTP/3 provides transport layer level stream multiplexing and accordingly it has encountered stream prioritization problem. The problem states the determination of which stream to transmit amongst multiple streams on a connection within limited network resources and this contributes to the completion time of web object loading. Meanwhile, dependency relationship between web activities exists and this implies that dependency relationship between web object loading activities also exists. In order to transfer web objects in accordance with the web page load process at the browser, we proposed a HTTP/3 stream prioritization scheme based on web object dependency. Particularly, we conducted the evaluation on a browser-based testbed we built rather than on HTTP/3 library. The proposed prioritization scheme was evaluated using the testbed and it was shown that the application of the scheme could improve the user’s experience.
Search
Journal of KIISE
- ISSN : 2383-630X(Print)
- ISSN : 2383-6296(Electronic)
- KCI Accredited Journal
Editorial Office
- Tel. +82-2-588-9240
- Fax. +82-2-521-1352
- E-mail. chwoo@kiise.or.kr