Earticle

최근 이미지 분류 작업의 기술들은 잘 정제되어 있는 대규모 데이터 세트에서 강력한 성능을 보여주지만, 긴꼬리 분 포에서는 약한 모습을 보인다. 또한, 계층적 데이터 세트에 대한 연구는 제한적이다. 그러나 트리 형태의 구조로 조 직된 계층적 데이터 세트에서는 상위 클래스 정보를 추가적으로 활용할 수 있다. 본 논문에서는 계층적 특징 활용을 위해 문화유산 데이터 세트를 수집하였으며, 해당 데이터 세트에서 분류 실패의 두 가지 원인을 발견하였다. 1) 클 래스 불균형으로 인한 오 분류, 2) 상위 클래스 오 분류. 우리는 이러한 문제를 해결하기 위해 두 가지의 손실 함수 를 활용한 학습 전략을 제안한다. 유니폼 샘플링과 클래스별 유니폼 샘플링, 두 가지 샘플링 방법의 이점을 모두 가 져가기 위한 consistency 손실 함수와 상위 클래스의 특성을 위해 하위 클래스 예측에 기반한 손실 함수를 사용하 여 문화유산 데이터 세트에서 베이스라인 대비 5.4%의 성능 향상을 달성한다.

Recent techniques in image classification show strong performance on large, well-refined datasets, but are weak on long-tailed distributions. In addition, there is limited research on hierarchical datasets; however, hierarchical datasets organized in a tree-like structure can further exploit higher-level class information. In this paper, we collected a cultural heritage dataset to utilize hierarchical features, and found two causes of classification failure in the dataset. 1) misclassification due to class imbalance, and 2) superclass misclassification. To solve these problems, we propose a learning strategy utilizing two loss functions. We use a consistency loss function to take advantage of both sampling methods, uniform sampling and class-specific uniform sampling, and a loss function based on subclass prediction for the characteristics of the parent class. We achieve a performance improvement of 5.4% over baseline on a cultural heritage dataset using our proposed learning technique.

대규모 글쓰기 평가에서 자동 에세이 채점(Automated Essay Scoring, AES)은 채점 효율을 높이지만, 생성형 LLM 기반 채점은 점수 토큰의 순서형 학습이 어렵고 동일 입력에서도 출력이 달라지는 변동성으로 채점 정확도, 신뢰도 문제가 발생할 수 있다. 본 연구는 추가 회귀 헤드 없이 단일 LLM이 루브릭별 점수 토큰과 피드백 텍스트를 하나의 autoregressive 생성 경로에서 순차 생성하는 통합 학습·추론 프레임워크를 제안하였다. 학습에서는 숫자 토큰 분포 기대값과 정답 점수의 오차를 최소화하는 숫자 토큰 손실(Number Token Loss, NTL)로 순서형 점수의 위계 구조를 반영하고, 피드백 생성의 표면형 일치 과의존을 완화하기 위해 정답 피드백 주변 의미 공간으로 예측 분포를 정렬하는 보조 목적함수 의미 정렬 손실(Semantic Alignment Loss, SAL)를 결합하였다. 추론에서는 자기 일관성(Self-Consistency)으로 다회 샘플을 생성한 뒤 평균 집계하여 확률적 변동성을 통계적으로 완화하였다. AI Hub '서술형 글쓰기 평가 데이터셋'에서 제안 방법은 평균 Quadratic Weighted Kappa (QWK) 0.626를 달성해 회귀헤드 기반 베이스라인(0.581)을 상회했다. 이는 NTL이 점수 예측 정밀도를 개선하고, 자기 일관성이 재현성과 강건성을 높이며, SAL이 점수 판정에 상보적 학습 신호를 제공함을 시사한다.

In large-scale writing assessments, Automated Essay Scoring (AES) enhances scoring efficiency. However, due to the difficulty in learning ordinality among score tokens and the high variance of outputs for identical prompts, LLM-based grading systems often struggle with scoring precision and consistency. This study proposes a unified training and inference framework where a single LLM sequentially generates rubric-specific score tokens and feedback text within a single autoregressive generation path, without the need for additional modules. During training, we incorporated Number Token Loss (NTL) to reflect the hierarchical structure of ordinal scores by minimizing the error between the expected value of the number token distribution and the ground truth. To mitigate over-reliance on surface-level matching during feedback generation, we also integrated Semantic Alignment Loss (SAL) as an auxiliary objective, which aligns the predicted distribution with the semantic space surrounding the ground truth feedback. During inference, stochastic variability was statistically mitigated by generating and averaging multiple samples through Self-Consistency. On the AI Hub 'Descriptive Writing Evaluation Dataset,' the proposed method achieved an Average Quadratic Weighted Kappa (QWK) of 0.626, outperforming regression-head-based baselines. These results suggest that NTL improves score prediction precision, Self-Consistency enhances reproducibility and robustness, and SAL provides complementary learning signals for score determination.

Semi-supervised learning (SSL) has shown great promise in utilizing both labeled and unlabeled data, especially using consistency regularization To enhance the model's learning efficiency, we proposed a dualbranch co-training approach. After dividing the training data into two subsets, each branch utilizes a teacherstudent model pair, where the teacher's weights are updated via an exponential moving average (EMA). The framework combines supervised loss and unsupervised loss to optimize the model. Upon label expansion (e.g pseudo labeling), an additional otherside view is introduced, promoting agreement between the branches' predictions on shared data. This loss mitigates errors arising from incorrect pseudo-labels and enhances the overall robustness of the training process. By dynamically adjusting pseudo-label inclusion based on confidence thresholds, our methodology reduces the impact of noisy data and prevents overfitting. As a result, we could demonstrate the effectiveness of the proposed method in leveraging unlabeled data while maintaining high performance.

합성곱 신경망 (Convolutional Neural Network, CNN)은 컴퓨터 비전 분야에서 활발히 적용되어 왔으며, 이미지 분류, 문서 분류, 지문 인식 등에서 탁월한 인식 능력을 보여 왔음을 여러 연구를 통해서 검증되었다. 본 연구는 시계열의 표면근전도 신호를 입력데이터로 취하는 숫자지화 인식 응용에 이미지 분류에서 탁월한 인식 성능을 보이는 합성곱 신경망을 적용한 것으로, 반복적인 한국 숫자지화 인식 수행에서도 일관된 학습을 수행하는지를 검증하는 연구로, 문헌에서 보기 힘든 연구이다. 이를 검증하기 위해, 한국 숫자지화 영(0)부터 다섯(5)까지의 여섯 숫자지화를 시연하도록 훈련한 실험 대상 1인의 아래팔 근육으로부터 획득한 숫자별 60개씩 총 360개의 표면근전도 신호를 획득하였으며, 그 중에서 252개의 표면근전도 신호를 입력데이터, 108개의 표면근전도 신호는 테스트데이터로 CNN 인식에 활용하였다. CNN 인식을 위해 필요한 학습단계는 100 학습단계, CNN 인식의 반복 수행 횟수는 10회로 설정하였으며, 반복 수행마다 테스트데이터를 활용하여 인식률을 계산하였다. 본 연구에서 실험한 결과에서 보듯이, 반복 인식마다 CNN의 학습은 일관되었으며, 99.1% 이상 (60 숫자지화 중 하나의 숫자지화 인식에 오류발생)의 높은 인식률을 보였다. 따라서, CNN 기법은 시계열의 표면근전도 신호를 입력데이터로 하는 숫자지화 인식 분야에서도 전역 솔루션과 함께 우수한 인식 능력을 제공하는 기법 중에 하나이다.

Convolutional Neural Network (CNN) has been actively employed in the application of computer vision, and has been proved to have its superior performance in image classification, document classification, and finger print recognition. This work focuses on an application of CNN, having outstanding performance in image classification, to recognition of korean finger number using time series sEMG signals as input and validates CNN's capability in providing its consistent learning in repeated application for recognition of sEMG based Korean finger numbers, which has been rarely a topic in previous studies. To this end, 252 sEMG signals as input data and 108 sEMG signals as test data out of 360 sEMG signals (60 signals each number) acquired from a forearm muscle of the subject who is trained to consistently perform six Korean finger number gestures from zero(0) to five(5) were used for CNN based finger number recognition. CNN was set to have 100 learning iterations for each application of finger number recognition, and to have 10 repetitive applications of finger number recognition for the consistency of CNN's learning. Recognition rate at each repetition was calculated from test data. As can be seen from the results in this work, CNN shows consistent learning at each repetitive application of finger number recognition and outstanding recognition rates of more than 99.1% (missed one case out of 60 cases). Thus, CNN is one of powerful techniques for finger number recognition based on time-series sEMG signals to provide not only global solution but also excellent recognition rates.

본 연구는 Bernas형 이온주입장치의 ion flux를 예측하기 위해, Gated Recurrent Unit(GRU) 및 Long Short-Term Memory(LSTM) 기반의 시계열 딥러닝 모델을 적용하였다. 입력 데이터는 장치 성능에 영향을 주는 이온원 구간의 14개 변수이며, 하이퍼 파라미터는 Tree-structured Parzen Estimator(TPE) 기반 Bayesian Optimization을 통해 탐색하였다. 그러나 TPE는 초기 조건에 민감해 학습 결과의 일관성이 떨어질 수 있어, 이를 보완하기 위해 Gradient Consistency를 통한 window size 선택 방식을 제안하였다. 성능 평가는 Mean Squared Error(MSE)와 Mean Absolute Error(MAE) 외에도 반복 실험 간 변동성을 측정하는 Relative Standard Deviation(RSD)를 활용하였다. 확인 결과, GRU 모델에서 validation MSE의 RSD가 TPE(70.3%) 대비 Gradient Consistency(GC) (57.0%)에서 감소하였고, test MAE 역시 35.5%에서 31.7%로 줄어들어 더 안정적인 성능을 보였다. 이를 통해 본 연구는 이온주입장치와 같은 고정밀 장비 운용에서, 평균 성능뿐만 아니라 일관성과 신뢰성 확보를 위한 새로운 기준으로 GC를 통한 window size 최적화 방안을 제시한다.

This study applied GRU and LSTM models to predict ion flux in Bernas-type ion implantation systems using 14 ion source variables. While TPE-based Bayesian Optimization explored hyperparameters, its sensitivity to initial conditions leads to inconsistent results. We proposed a Gradient Consistency-based window size selection to address this limitation. Performance evaluation using MSE, MAE, and RSD showed that the GRU model achieved better stability: validation MSE RSD decreased from 70.3% (TPE) to 57.0% (Gradient Consistency-based), and test MAE improved from 35.5% to 31.7%. This approach provides a new criterion for ensuring consistency in high-precision equipment operation.

The purpose of this paper is to investigatethe relationship between Korean elementary teachers' teaching practice and students' perceptions of classroom learning environment. It is important to examine whether teachers conceive of mathematics and their teaching practice are consistent with their actual teaching mathematics in classroom. To do so, two reform-oriented and two traditionally-oriented elementary teachers in Seoul and a total of 130 students from these four selected teachers' classes participated in the present study Using classroom environment survey scale [CES] the differences in students' perceptions on classroom learning environment were examined. The results show that students taught by reform-oriented teachers have significantly higher score in the area of participation, relevance, and commitment to learning than traditionally-oriented group of students. Thus, we may conclude that the reform-oriented group of students have more benefits in learning mathematics than do the traditionally-oriented group of students.

The purpose of this study was to verify the effect on learning using YouTube. The relationship between expectations, cognitive attitudes, emotional attitudes, and reuse intentions was explored. The results are as follows. First, the factors affecting the intention to reuse YouTube were emotional attitude and expectation. cognitive attitude did not show any significant effect. Second, cognitive attitude and expectation agreement were significant variables in emotional attitude. cognitive attitude was identified as a major factor affecting emotional attitude. Third, expectation agreement was found to have a significant influence on cognitive attitude. Therefore, expectancy is an important factor in identifying cognitive attitudes. In particular, the satisfaction of expectations after experiencing a new technology or system, such as YouTube, affects the cognitive attitude. Finally, the main findings of this study were that cognitive attitude was not a variable affecting the intention to reuse YouTube. The reason for this is that YouTube is used as a medium of interest, and it is not used as a medium for searching the main data source for learning. Therefore, in order to activate YouTube learning, it is necessary to convert recognition into YouTube for learning rather than YouTube for play.