Digital Library[ Search Result ]
A Quantitative Comparison of LIME and SHAP using Stamp-Based Distance Method on Image Data
http://doi.org/10.5626/JOK.2023.50.10.906
XAI(eXplainable AI), 인공신경망, MNIST, 도장 기반의 distance method, LIME, SHAP Abstract XAI, or eXplainable AI, is a technique used to explain artificial neural networks in a way that can be understood by humans. However, it is difficult to compare explanations and heat maps produced by XAI algorithms numerically as it is unclear how humans interpret them. This presents a challenge in determining which XAI algorithm is the most effective and accurate in providing explanations. Therefore, we introduced a stamp-based distance method to compare several XAI algorithms and identify the most accurate algorithm. The proposed method involves evaluating the quality of explanations generated by XAI algorithms applied to a deep learning model trained to detect the presence of stamps in the MNIST dataset. This evaluation was performed using statistical techniques to determine the effectiveness of each XAI algorithm. This paper evaluated performances of LIME and SHAP algorithms using the distance method, which compared explanations produced by each algorithm. Result revealed that LIME with the Felzenszwalb method provided more effective explanations than other LIME and SHAP algorithms.
Document-level Machine Translation Data Augmentation Using a Cluster Algorithm and NSP
http://doi.org/10.5626/JOK.2023.50.5.401
In recent years, research on document level machine translation has been actively conducted to understand the context of the entire document and perform natural translation. Similar to the sentence-level machine translation model, a large amount of training data is required for training of the document-level machine translation model, but there is great difficulty in building a large amount of document-level parallel corpus. Therefore, in this paper, we propose a data augmentation technique effective for document-level machine translation in order to improve the lack of parallel corpus per document. As a result of the experiment, by applying the data augmentation technique using the cluster algorithm and NSP to the sentence unit parallel corpus without context, the performance of the document-level machine translation is improved by S-BLEU 3.0 and D-BLEU 2.7 compared to that before application of the data augmentation technique.
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.
Hyperbolic Graph Transformer Networks for non-Euclidean Data Analysis on Heterogeneous Graphs
Seunghun Lee, Hyeonjin Park, Hyunwoo J Kim
http://doi.org/10.5626/JOK.2021.48.2.217
Convolution Neural Networks (CNNs), which are based on convolution operations, are used for various tasks in image classification, image generation, time series analysis, etc. Since the convolution operations are not directly applicable to non-Euclidean spaces such as graphs and manifolds, a variety of Graph Neural Networks (GNNs) have extended convolutional neural networks to homogeneous graphs, which has a single type of edges and nodes. However, in real-world applications, heterogeneous and hierarchical graph data often occur. To expand the operating range of GNNs to the graphs that have multiple types of nodes and edges with the hierarchy, herein, we propose a new model that integrates Hyperbolic Graph Convolution Networks (HGCNs) and Graph Transformer Networks (GTNs).
Morpheme-based Efficient Korean Word Embedding
Dongjun Lee, Yubin Lim, Ted “Taekyoung” Kwon
http://doi.org/10.5626/JOK.2018.45.5.444
Previous word embedding models such as word2vec and Glove are not able to learn the internal structure of words. This is a serious limitation for agglutinative languages with morphology such as Korean. In this paper, we propose a new model which is an expansion of the previous skip-gram model. This defines each word vector as a sum of its morpheme vectors and hence, learns the vectors of morphemes. To test the efficiency of our embedding, we conducted a word similarity test and a word analogy test. Furthermore, using our trained vectors on other NLP tasks, we tested how much performance actually had been enhanced.
End-to-end Korean Document Summarization using Copy Mechanism and Input-feeding
In this paper, the copy mechanism and input feeding are applied to recurrent neural network(RNN)-search model in a Korean-document summarization in an end-to-end manner. In addition, the performances of the document summarizations are compared according to the model and the tokenization format; accordingly, the syllable-unit, morpheme-unit, and hybrid-unit tokenization formats are compared. For the experiments, Internet newspaper articles were collected to construct a Korean-document summary data set (train set: 30291 documents; development set: 3786 documents; test set: 3705 documents). When the format was tokenized as the morpheme-unit, the models with the input feeding and the copy mechanism showed the highest performances of ROUGE-1 35.92, ROUGE-215.37, and ROUGE-L 29.45.
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