Digital Library[ Search Result ]
Explainable Graph Neural Network for Medical Science Research
Yewon Shin, Kisung Moon, Youngsuk Jung, Sunyoung Kwon
http://doi.org/10.5626/JOK.2022.49.11.999
Explainable AI (XAI) is a technology that provides explainability for our end-users to comprehend prediction results of ML algorithms. In particular, the reliability of the decision-making process of an AI algorithm through XAI technology is the most critical in the medical field in terms of real applications. However, complex interaction-based medical data restrict the application of existing XAI technologies developed mostly for image or text data. Graph Neural Network (GNN)-based XAI research has been highlighted in recent years because GNN is technically specialized to capture complex relationships in data. In this paper, we proposed a taxonomy according to the application method and algorithm of GNN-based XAI technology with current XAI research trends and its use-cases in four detailed areas of the medical field. We also expounded on the technical limitations and future works of XAI research specialized in the biomedical area.
Automatic Classification of Pneumonia Based on Ensemble Deep Learning Model Using Intensity Normalization and Multiscale Lung-Focused Patches on Chest X-Ray Images
Yoon Jo Kim, Jinseo An, Helen Hong
http://doi.org/10.5626/JOK.2022.49.9.677
It is difficult to classify normal and pneumonia in pediatric chest X-ray (CXR) images due to irregular intensity values. In addition, deep learning model has a limitation in that it can misclassify CXR by incorrectly focusing on the outer part of the lung. This study proposed an automatic classification of pneumonia based on ensemble deep learning model using three intensity normalizations and multiscale lung-focused patches on CXR images. First, to correct for irregular intensity values in internal lungs, three intensity normalization methods were performed respectively. Second, to focus on internal lungs, regions of interest were extracted by segmenting lung regions. Third, multiscale lung-focused patches were extracted to train the characterization of pneumonia. Finally, ensemble modeling with attention module was performed to improve the classification performance. In the experiment, the method using large patches of CLAHE images showed an accuracy of 92%, which was 5% higher than that of original images. Furthermore, the proposed method using an ensemble of large and middle patches showed the best performance with an accuracy of 93%.
Knowledge Graph Embedding for Link Prediction using Node-Link Interaction-based Graph Attention Networks
http://doi.org/10.5626/JOK.2022.49.7.555
Knowledge graphs are structures that express knowledge in the real world in the form of nodes and links-based triple form. These knowledge graphs are incomplete and many embedding techniques have been studied to effectively represent nodes and links in low-dimensional vector spaces to find other missing relationships. Recently, many neural network-based knowledge graph link prediction methods have been studied. However existing models consider nodes and links independently when determining the importance of a triple to a node which makes it difficult to reflect the interaction between nodes and links. In this paper, we propose an embedding method that will be used to analyze the importance of triple units by simultaneously considering nodes and links using composition operators, and at the same time prove that the model outperforms other methods in knowledge graph link prediction.
Efficient Compilation Error Localization with DNN
http://doi.org/10.5626/JOK.2022.49.6.434
There are few programs with no compilation errors. The compiler provides the programmers with compiler error messages as clues to solve the problem, but analyzing the error messages correctly also consumes much time. Although there are many proposals that suggest the error localization method and how to repair the error, most of the proposals are using data from novice programmers, or can be applied only to one specific programming language. It is difficult to apply practically in large-scale projects conducted in the company. In this study, to increase the efficiency of compile error handling in practical projects, we propose DeepErrorFinder which identifies the location of compilation errors using DNN. This model, which is based on the LSTM model, predicts the error location after training based on compilation error logs, and repair changes from mobile phone software development projects. As a result of the experiments, it showed an accuracy of 52% and reduced the elapsed time compared to a manual search. It can facilitate quickly finding the location of the compilation error code in practice projects.
Current Research Trends in Attacking Deep Learning Using Adversarial Examples
http://doi.org/10.5626/JOK.2022.49.5.339
Despite its great success, deep learning is largely vulnerable to attacks by adversarial examples. An attack by an adversarial example is a technique that deceives a well-trained deep learning model by adding noise to the normal input that is small enough to be invisible to the human eye. Such vulnerabilities in safe-critical systems, such as autonomous cars, can cause catastrophic failures and results such as traffic accidents. Although such attacks using adversarial examples with high potential risk are being actively studied worldwide, there is still a lack of systematic summaries on this in the Korean academic community. Therefore, in this paper, we systematically summarized recent trends in adversarial attacks on deep learning to boost future research.
SVD-based Cross-Domain Recommendation Using K-means Clustering
http://doi.org/10.5626/JOK.2022.49.5.360
Cross-domain recommendation is a method that shares related user information data and item data in different domains. It is mainly used in online shopping malls with many users or multimedia service contents, such as YouTube or Netflix. Through K-means Clustering, embeddings are created by performing clustering based on user data and ratings. After learning the result through a multi-layer neural network, user satisfaction is predicted. Then, items suitable for the user are recommended using matrix factorization, which is a collaborative filtering technique. Through this study, it was shown through experiments that recommendations can predict cold-start problems at a lesser time cost and increase the user satisfaction.
Detecting Design Infringement Using Multi-Modal Visual Data and Auto Encoder based on Convolutional Neural Network
Jeonggeol Kim, Jiyou Seo, Chanjae Lee, Seongmin Jo, Seungmin Kim, Seokmin Yoon, Young Yoon
http://doi.org/10.5626/JOK.2022.49.2.137
Recently, it has become very difficult to distinguish between counterfeit products and authentic goods, and the volume of these forgeries is increasing at an alarming rate. Prompt detection of these counterfeit products is challenging since only humans can identify these forgeries through trained expertise. In this paper, given the photograph and design drawing, we use convolutional neural networks and auto-encoders to detect the possible infringement of design rights without dissembling or damaging the suspected items. We have developed an easy-to-expand system that supports the constant addition of new goods to be examined. We present the result of our system tested with a set of authentic and forged goods.
Estimation of Finger Motion using Transient EMG Signals
http://doi.org/10.5626/JOK.2022.49.2.157
In this paper, we propose a deep learning model for estimating finger movements based on EMG signals. We have also evaluated and analyzed the accuracy of the model. We have applied the U-Net structure, which is widely used in medical image analysis, to our model. In general, U-Net is mainly used for processing of two-dimensional images. However, in this paper, 8-channel one-dimensional time series EMG data is used as inputs, and information about finger movement is obtained as results. We have acquired the data set consisting of 8,000 motions, which is divided into the training and evaluation data sets. The accuracy of the prediction of our model is about 89.32%.
Improving False Positive Rate of Extended Learned Bloom Filters Using Grid Search
http://doi.org/10.5626/JOK.2022.49.1.78
Bloom filter is a data structure that represents a set and returns whether data is included or not. However, there are cases in which false positives are returned at the cost of using less space. The learned bloom filter is a variation of the bloom filter, that uses a machine learning model in the pre-processing process to improve the false-positive rate. The learned bloom filter stores some data in the machine learning model, and the leftover data is stored in the auxiliary filter. An auxiliary filter can be implemented by using a bloom filter only, but in this paper, we use the bloom filter and the learned hash function, and this is called an extended learned bloom filter. The learned hash function uses the output value of the machine learning model as a hash function. In this paper, we propose a method that improves the false positive rate of the extended learned bloom filter through grid search. This method explores the extended learned bloom filter with the lowest false positive rate, by increasing the hyperparameter that represents the ratio of the learned hash function. As a result, we experimentally show that the extended learned bloom filter selected through grid search, can have a 20% improvement in false-positive rate compared to the learned bloom filter, in the experiment that needs more than 100,000 data to store. In addition, we also show that the false negative error may occur in the learned hash function by the use of 32-bit floating points in the neural network model. This can be solved by changing the floating points to 64-bit. Finally, we show that in an experiment where we query 10,000 data, we can adjust the structure of the neural network model to save 20KB of space and create an extended learned bloom filter with the same false-positive rate. However, the query time is increased by 2% at the cost of saving 20KB of space.
Deletion-based Korean Sentence Compression using Graph Neural Networks
Gyoung-Ho Lee, Yo-Han Park, Kong Joo Lee
http://doi.org/10.5626/JOK.2022.49.1.32
Automatic sentence compression aims at generating a concise sentence from a lengthy source sentence. Most common approaches to sentence compression is deletion-based compression. In this paper, we implement deletion-based sentence compression systems based on a binary classifier and long short-term memory (LSTM) networks with attention layers. The binary classifier, which is a baseline model, classifies words in a sentence into words that need to be deleted and words that will remain in a compressed sentence. We also introduce a graph neural network (GNN) in order to employ dependency tree structures when compressing a sentence. A dependency tree is encoded by a graph convolutional network (GCN), one of the most common GNNs, and every node in the encoded tree is input into the sentence compression module. As a conventional GCN deals with only undirected graphs, we propose a directed graph convolutional network (D-GCN) to differentiate between parent and child nodes of a dependency tree in sentence compression. Experimental results show that the baseline model is improved in terms of the sentence compression accuracy when employing a GNN. Regarding the performance comparison of graph networks, a D-GCN achieves higher F1 scores than a GCN when applied to sentence compression. Through experiments, it is confirmed that better performance can be achieved for sentence compression when the dependency syntax tree structure is explicitly reflected.
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