• Journal of Intelligence and Information Systems
• /
• v.29 no.3
• /
• pp.125-144
• /
• 2023

Recently, the use of Large Language Models (LLMs) such as ChatGPT has been increasing in various fields such as interactive commerce and mobile financial services. However, LMMs, which are mainly created by learning existing documents, can also learn various human biases inherent in documents. Nevertheless, there have been few comparative studies on the aspects of bias and discrimination in LLMs. The purpose of this study is to examine the existence and extent of nine types of discrimination (Age, Disability status, Gender identity, Nationality, Physical appearance, Race ethnicity, Religion, Socio-economic status, Sexual orientation) in LLMs and suggest ways to improve them. For this purpose, we utilized BBQ (Bias Benchmark for QA), a tool for identifying discrimination, to compare three large-scale language models including ChatGPT, GPT-3, and Bing Chat. As a result of the evaluation, a large number of discriminatory responses were observed in the mega-language models, and the patterns differed depending on the mega-language model. In particular, problems were exposed in elder discrimination and disability discrimination, which are not traditional AI ethics issues such as sexism, racism, and economic inequality, and a new perspective on AI ethics was found. Based on the results of the comparison, this paper describes how to improve and develop large-scale language models in the future.

• Journal of the Institute of Electronics and Information Engineers
• /
• v.50 no.9
• /
• pp.85-91
• /
• 2013

This paper proposes the data reusable search scan methods for full search and fast search to implement low power Motion Estimation (ME). The proposed Optimized Sub-region Partitioning (OSP) method which divide search region into several sub-region can reduce the number of the required Reconfigurable Register Array (RRA) by half compared to the existing smart snake scan method for the same data reusability. In addition, the proposed Center Biased Search Scan method (CBSS) for various fast search algorithms can improve the data reusability. The performance comparisons show that the proposed search scan methods can reduce the average redundant data loading about 26.9% and 16.1% compared with the existing rater scan and snake scan methods, respectively. Due to the reduction of memory accesses, the proposed search scan methods are quite suitable for low power and high performance ME implementation.

• KIPS Transactions on Software and Data Engineering
• /
• v.9 no.9
• /
• pp.281-290
• /
• 2020

The Vision-and-Language Navigation(VLN) task is a complex intelligence problem that requires both visual and language comprehension skills. In this paper, we propose a new learning model for visual-language navigation agents. The model adopts a hybrid learning that combines imitation learning based on demo data and reinforcement learning based on action reward. Therefore, this model can meet both problems of imitation learning that can be biased to the demo data and reinforcement learning with relatively low data efficiency. In addition, the proposed model uses a novel path-based reward function designed to solve the problem of existing goal-based reward functions. In this paper, we demonstrate the high performance of the proposed model through various experiments using both Matterport3D simulation environment and R2R benchmark dataset.

• KIPS Transactions on Computer and Communication Systems
• /
• v.11 no.12
• /
• pp.445-452
• /
• 2022

Class distribution of unbalanced data is an important part of the digital world and is a significant part of cybersecurity. Abnormal activity of unbalanced data should be found and problems solved. Although a system capable of tracking patterns in all transactions is needed, machine learning with disproportionate data, which typically has abnormal patterns, can ignore and degrade performance for minority layers, and predictive models can be inaccurately biased. In this paper, we predict target variables and improve accuracy by combining estimates using Synthetic Minority Oversampling Technique (SMOTE) and Light GBM algorithms as an approach to address unbalanced datasets. Experimental results were compared with logistic regression, decision tree, KNN, Random Forest, and XGBoost algorithms. The performance was similar in accuracy and reproduction rate, but in precision, two algorithms performed at Random Forest 80.76% and Light GBM 97.16%, and in F1-score, Random Forest 84.67% and Light GBM 91.96%. As a result of this experiment, it was confirmed that Light GBM's performance was similar without deviation or improved by up to 16% compared to five algorithms.

• Journal of Korea Society of Industrial Information Systems
• /
• v.29 no.2
• /
• pp.15-25
• /
• 2024

For effective analysis of animal ecosystems, technology that can automatically identify the current status of animal habitats is crucial. Specifically, animal sound classification, which identifies species based on their sounds, is gaining great attention where video-based discrimination is impractical. Traditional studies have relied on a single deep learning model to classify animal sounds. However, sounds collected in outdoor settings often include substantial background noise, complicating the task for a single model. In addition, data imbalance among species may lead to biased model training. To address these challenges, in this paper, we propose an animal sound classification scheme that combines predictions from multiple models using Focal Loss, which adjusts penalties based on class data volume. Experiments on public datasets have demonstrated that our scheme can improve recall by up to 22.6% compared to an average of single models.

• Annual Conference on Human and Language Technology
• /
• 2018.10a
• /
• pp.291-295
• /
• 2018

한국어는 교착어적 특성이 강한 언어로, 교착어적 특성이 없는 영어 등의 언어와 달리 형태소의 수에 따라 조합 가능한 어절의 수가 매우 많으므로 어절 단위의 처리가 매우 어렵다. 따라서 어절을 더 작은 단위로 분해하는 전처리 단계가 요구되는데, 형태소 분석이 이를 위해 주로 사용되었다. 하지만 지도학습 방법을 이용한 형태소 분석 시스템은 다량의 학습 데이터가 요구되고, 비지도학습 방법을 이용한 형태소 분석은 성능에 큰 하락을 보인다. Byte Pair Encoding은 데이터를 압축하는 알고리즘으로, 이를 자연어처리 분야에 응용하면 비지도학습 방법으로 어절을 더 작은 단위로 분해할 수 있다. 본 연구에서는 한국어에 Byte Pair Encoding을 적용하는 두 가지 방법인 음절 단위 처리와 자모 단위 처리의 성능 및 특성을 정량적, 정성적으로 분석하는 방법을 제안하였다. 또한, 이 방법을 세종 말뭉치에 적용하여 각각의 알고리즘을 이용한 어절 분해를 실험하고, 그 결과를 어절 분해 정확도, 편향, 편차를 바탕으로 비교, 분석하였다.

• Proceedings of the Korea Water Resources Association Conference
• /
• 2021.06a
• /
• pp.393-393
• /
• 2021

도시화는 도시 호우 유출 발생으로 인한 수질 악화를 초래했고 문제를 해결하기 위해 본 연구에서는 보다 정확한 설계를 위해 그린인프라(Green Infrastructure, GI)의 구조적 특성과 수문학적인 특성을 이용해 어떤 인자들이 설계에 필요한지 상관관계를 통해 분석하였다. GI의 종류 중 저류지와 저류연못의 총부유사량(Total Suspended Solids, TSS)와 총인 (Total Phosphorous, TP)의 유입수, 유출수, 비점오염원 농도, 수문학적인 특성 그리고 GI의 구조적 특성을 Ordinary Least Squares regression(OLS)과 Multi Linear Regression(MLR) 방법을 적용하였다. GI의 구조적인 특성은 한 BMP마다 달라지지 않으나 호우사상의 데이터 개수에 의한 편향이 있을 수 있다. 이런 문제를 해결하기 위해 일정한 범위를 가지고 무작위로 데이터를 추출하는 방법과 이상치를 제외하는 방법을 사용하여 모델에 적용하였다. 이러한 OLS와 MLR 모델들의 정확도를 PBIAS(Percent Bias), NSE(Nash-Sutcliffe efficiency), RSR(RMSE-observations standard deviation ratio)을 통해 분석할 수 있다. 연구 결과 유입수의 비점오염원의 농도뿐만 아니라 수문학적 특성과 GI의 구조적 특성이 함께 들어갈 시 더 좋은 상관관계를 가지고 있음을 알 수 있다. 저류지가 저류연못보다 모델의 성능평가 면에서 좋은 값을 가지고 있지만 특성별 상관관계는 저류연못이 더 뚜렷한 결과를 보여준다.

• Journal of the Korean Society for information Management
• /
• v.19 no.4
• /
• pp.35-51
• /
• 2002

Inductive learning and classification techniques have been employed in various research and applications that organize textual data to solve the problem of information access. In this study, we develop hybrid model combination methods which incorporate the concepts and techniques for multiple modeling algorithms to improve the accuracy of text classification, and conduct experiments to evaluate the performances of proposed schemes. Boosted stacking, one of the extended stacking schemes proposed in this study yields higher accuracy relative to the conventional model combination methods and single classifiers.

• Journal of the Korean Geophysical Society
• /
• v.4 no.3
• /
• pp.175-182
• /
• 2001

We consider the problem of determining smoothing parameters of Gibbs priors for Bayesian methods used in the medical imaging application of emission tomographic reconstruction. We address a simple smoothing prior (membrane) whose global hyperparameter (the smoothing parameter) controls the bias/variance tradeoff of the solution. We base our maximum-likelihood (ML) estimates of hyperparameters on observed training data, and argue the motivation for this approach. Good results are obtained with a simple ML estimate of the smoothing parameter for the membrane prior.

• Proceedings of the Korean Information Science Society Conference
• /
• 2004.10a
• /
• pp.226-228
• /
• 2004

베이지안 네트워크는 불확실한 상황을 모델링하기 위한 확률 기반의 모델이다. 베이지안 네트워크의 구조를 자동 학습하기 위한 연구가 많이 있었고, 최근에는 진화 알고리즘을 이용한 연구가 많이 진행되고 있다. 그러나 대부분은 마지막 세대의 가장 좋은 개체만을 이용하고 있다. 시스템이 요구하는 다양한 요구조건을 하나의 적합도 평가 수식으로 나타내기 어렵기 때문에, 마지막 세대의 가장 좋은 개체는 종종 편향되거나 변화하는 환경에 덜 적응적일 수 있다. 본 논문에서는 적합도 공유 방법으로 다양한 베이지안 네트워크를 생성하고, 이를 베이즈 규칙을 통해 결합하여 변화하는 환경에 적응적인 추론 모델을 구축할 수 있는 방법을 제안한다. 성능 평가를 위해 ALARM 네트워크에서 인공적으로 생성한 데이터를 이용한 구조 학습 및 추론 실험을 수행하였다. 다양한 조건에서 학습된 네트워크를 실험한 결과, 제안한 방법이 변화하는 환경에서 더욱 강건하고 적응적인 모델을 생성할 수 있음을 확인한 수 있었다.