Earticle

최근 딥러닝 기반 음성 인식 기술의 발전으로 사용자 발화를 이용한 서비스 개발이 활발하다. 하지만 음성 인식 기술을 조선소에 적용하기 위해서 현장 소음의 다양성과 불규칙성, 통신 장애, 하드웨어 비용 등의 제약조건이 존재한다. 또한, 다수의 철 구조물이 존재하고 통신 음영구역에서 작업이 진행되는 조선업 특성상 서버에서 동작하는 대규모 음성 인식 모델이 아닌 작업자 장비에 부착된 소형 모듈에 탑재하여 독립적으로 동작하는 소형 음성 인식 모델에 관한 연구가 필요하다. 본 연구에서는 조선소에서 활용할 수 있는 소형 음성 인식 모델을 학습하기 위한 Part-mix 증강기법과 NKB(Noise-Keyword Balancing) 손실함수라는 두 가지 기법을 제시한다. 이를 통해, 작업 소음과 작업자 간 대화가 섞여 발생하는 부분음으로 인한 오인식 문제를 극복하고 다수 소음 데이터, 소수 키워드 데이터의 불균형으로 발생하는 성능 저하 문제를 데이터 분포에 따른 Loss 조정으로 이를 개선한다.

Recent advancements in deep learning-based speech recognition technology have led to the development of services using user speech. However, applying this technology to shipyards involves overcoming challenges such as diverse and irregular noise, communication interruptions, and hardware costs. Due to the shipbuilding environment, with numerous steel structures and communication shadow zones, there is a need for small, standalone speech recognition models mounted on worker equipment, rather than large-scale server-based models. This study proposes two techniques—Part-mix augmentation and the NKB (Noise-Keyword Balancing) loss function—that jointly address misrecognition caused by mixed operational noise and partial speech from conversations. Together, these techniques overcome challenges arising from data imbalance between abundant noise and limited keyword data by adjusting the loss distribution based on data properties, significantly improving speech recognition performance in complex noise environments.

본 연구에서는 도로 인프라가 교차로에서 회전하는 차량의 궤적을 예측하기 위해 딥러닝 아키텍처인 트랜스포머 기반의 모형을 효과적으로 이용하는 방법론을 제안하고자 한다. 회전 구간은 차량의 진행 방향이 급격히 변화하는 경우가 많아 직진 구간에 비해 예측이 어렵기 때 문에 더 정밀한 모델이 필요하다. 회전 구간에 특화된 예측 모형을 위해 미시 교통 시뮬레이션 을 이용해 학습 데이터를 생성하고 모형을 학습시켰다. 두 가지 실험을 통해 손실함수와 입력 시퀀스 길이가 궤적 예측에 미치는 영향을 분석했으며, 회전 구간에서의 예측 성능 향상을 위 해서는 Smooth L1 loss 손실함수와 1초 내외의 입력 시퀀스를 사용하는 모델이 가장 우수한 성능을 보였다. 이 연구의 결과는 자율주행 차량의 경로 계획 및 교통 관리 시스템의 효율성 향상에 기여할 것으로 기대된다.

In this study, we propose an effective methodology based on the well-known transformer deep learning architecture to predict the trajectories of vehicles turning at intersections. From the road infrastructure perspective, segments with abrupt changes in a vehicle's direction are more difficult to predict than straight segments, so a more precise model is required. We generated training data using a microscopic traffic simulation, and trained the transformer-based prediction model for turning trajectories. Through two experiments, we analyzed the impact of loss functions and input sequence lengths on trajectory prediction. The results showed that models using a smooth L1 loss function and input sequences of approximately one second exhibited the best performance, improving prediction accuracy. The findings of this research are expected to enhance the efficiency of path planning for autonomous vehicles and traffic management systems.

Robust design is an approach to reduce the performance variation of mutiple responses in products and processes. In fact, in many experimental designs require the simultaneous optimization of multiple responses. In this paper, we propose how to simultaneously optimize multiple responses for robust design when data are collected from a combined array. The proposed method is based on the quadratic loss function. An example is illustrated to show the proposed method.

This paper is to introduce attribute acceptance sampling plan based on statistical inference of binomial proportions such as PPM or PPB. In addition, this paper presents three variable sampling acceptance sampling plans based on Cpm, Cpmk, and Taguchi's loss function. Producers are able to consider as not only external vendors but also internal customers.

Anterior spinal artery syndrome refers to the paralysis of the bilateral upper extremities, bladder dysfunction and the sensory deficit of pain and temperature below the level of injury. A 64 year-old female got a cardiac arrest event after stent insertion into the coronary artery. After CPR, she underwent the motor deficit (Z-T) of the bilateral upper extremities without any sensory deficit; proprioception, vibration and pain. The brain MRI showed no abnormality, but high signal intensity was detected in C3-C7 level by T2 sagittal plane and at the anterior horn area of gray matter by axial view of spine MRI. The ventilator has been applied after CPR. By the fluoroscopy, the movement of the diaphragm was decreased, and the nerve conduction study of both phrenic nerves showed no responses.

Brassinosteroids (BRs) are plant steroid hormones that play essential roles in growth and development. Mutations in BR-signaling pathways cause defective in growth and development like dwarfism, male sterility, abnormal vascular development and photomorphogenesis. Transition from vegetative to reproductive growth is a critical phase change in the development of a flowering plant. In a screen of activation-tagged Arabidopsis, we identified a mutant named abz126 that displayed longer hypocotyls when grown in the dark on MS media containing brassinazole (Brz), an inhibitor of BRs biosynthesis. We have cloned the mutant locus using adapter ligation PCR walking and identified that a single T-DNA had been integrated into the ninth exon of the GIGANTEA (GI) gene, involved in controling flowering time. This insertion resulted in loss-of-function of the GI gene and caused the following phenotypes: long petioles, tall plant height, many rosette leaves and late flowering. RT-PCR assays on abz126 mutant showed that the T-DNA insertion in GIGANTEA led to the loss of mRNA expression of the GI gene. In the hormone dose response assay, abz126 mutant showed: 1) an insensitivity to paclobutrazole (PAC), 2) an altered response with 6-benzylaminopurine (BAP) and 3) insensitive to Brassinolide (BL). Based on these results, we propose that the late flowering and tall phenotypes displayed by the abz126 mutant are caused by a loss-of-function of the GI gene associated with brassinosteroid hormone signaling.

This study was conducted on elite bodybuilders for 6 weeks using two weight loss methods (traditional method=high-weight loss vs. new method=low-weight loss). The subjects of this study were 26 male ac-tive elite bodybuilders who voluntarily wanted to participate in this ex-periment, and were divided into experimental group and traditional group. The traditional and experimental groups underwent dietary re-strictions and training for 6 weeks. The conclusion obtained from the results of this study is that in the case of anthropometric variables, there is a difference in weight class between the two groups. The ex-perimental group’s upper arm circumference was maintained com-pared to the pre-measured value after post-measurement, while the traditional group decreased, and a significant interaction effect was observed. In particular, a significant interaction effect was observed. In that the absolute value of maximum oxygen intake was significantly in-creased only in the post-experimental group compared to the pretest, and an interaction effect was observed, the use of low-weight loss method is more effective than the traditional method of high-weight loss method. An increase in some cytokines was observed despite tradition-al weight loss, but it did not cause an excessive decrease in immunity or a sharp decrease in performance factors. As a result of this experi-ment, it is judged that the use of a low-weight loss method is more pref-erable than the traditional high-weight loss method in relation to perfor-mance variables.

이상 행동 탐지는 비정상적인 행동을 취하는 물체를 추적하고 관찰하여 모니터링 또는 감독이 필요 한 분야에서 사용된다. 사람의 지속적인 개입없이 인공지능 알고리즘을 적용하여 자동화된 탐지 시스 템을 구현하기 위해서 광학 흐름 기법을 활용한 시공간적 특이도를 관찰하는 방법이 활용되고 있으며, 정해진 표현 범위가 없는 수많은 이상 행동을 식별하기에는 비지도학습 방법이 적합하기 때문에, 최근 생성적 적대 신경망을 통해 카메라에서 촬영된 동영상 프레임의 시공간적 특징을 분석하는 딥러닝 네 트워크에 대한 연구가 진행되고 있다. 이 중 영상 변환 네트워크는 복원 오류를 줄이기 위해 GAN 손 실과 L1 손실 함수를 최소화하도록 최적화되어 있다. 본 연구에서는 네트워크의 목적 함수를 조정하 여 모델의 정확도를 높이기 위한 실험을 수행하였다. 특히 구조적 유사도를 비교하여 손실을 최소화하 는 SSIM 손실 함수는 두 영상의 유사도를 측정하는 과정에서 많이 활용되고 있으며, 이 함수는 알고 리즘을 통해 계산된 광학 흐름 영상과 생성적 적대 신경망 모델의 생성자가 만들어낸 가짜 광학 흐름 영상의 차이를 계산하는 데 적합하므로 이를 추가하여 학습 후 프레임 수준 성능에서 더 높은 정확도 를 보임을 확인하였다. 그리고 네트워크의 목적 함수를 정의하는 과정에서 SSIM 손실 함수가 모델 훈련의 최적화에 어떻게 기여하는지 분석했다.

본 논문에서는 추가적인 데이터의 수집 없이 STS(Semantic Textual Similarity) task에서 multiple loss와 label smoothing 기법을 활용하여 성능을 향상시키는 방법을 제안한다. 기존의 모델들은 3-way softmax loss또는 cosine-similarity with MSE loss를 활용하여 STS task에 대한 손실 함숫값으로 계산하기 때문에, 사전 훈련된 언어 모델의 Word Contextual Embedding Vector가 좋지 않은 방향으로 변할 수 있는 문제점이 있다. 따라서 softmax loss에 label Smoothing 기법을 활용하여 유사한 문장끼리 클러스터링이 되도록 학습을 하였으며, 사전 언어 모델 중의 하나인 GPT(Generative Pre-training)에서 사전 훈련을 할 때 이용했던 loss 계산 방식에서 아이디어를 얻어 새로운 loss를 제안하는 모델에 결합하여 활용하였다. 이 2가지 접근 방식은 fine-tuning시에 활용하여 성능을 향상시킨다. 사전 훈련에는 Multi-genre Natural Language Inference, Stanford Natural Language Inference 데이터를 이용하였으며, 모델을 평가하기 위한 데이터로는 Semantic Textual Similarity benchmark를 이용하였다. 평균적으로 1.24% 정도의 성능 향상을 보였으며, Spearman’s rank에서는 최대 4.67%, Pearson correlation coefficient에서는 최대 5.64%의 성능 향상을 보였다.

Process capability indices are widely used in industries and quality assurance system. When designing the parameter on the multiple quality characteristics, there has been a study for optimization of problems, but there has been few former study on the possible conflicting phenomena in considertion of the correlations among the characteristics. To solve the issue on the optimal design for muliple quality characteristics, the study propose the expected loss function with cross-product terms among the characteristics and derived range of the coefficients of terms. Therefore, the analysis have to be required a multivariate statistical technique. This paper introduces to multivariate capability indices and then selects a multivariate process capability index incorporated both the process variation and the process deviation from target among these indices under the multivariate normal distribution. We propose a new multivariate capability index MCpm using quality loss function instead of the process variation and this index is compared with the proposed indices when quality characteristics are independent and dependent of each other,

We purpose a decision model to select the optimal facilities for the Decision Making problems with multiple characteristics(nominal-is-best characteristics, larger-is -better characteristics, smaller- is -better characteristics). Using this model, concept of the loss function is used in this comprehensive method of for select the optimal preferred facilities. To solve the issue on the optimal preferred facilities for multiple characteristics, this study propose the loss function with cross-product terms among the characteristics and derived range of the coefficients of the terms.

Process capability indices are widely used in industries and quality assurance system. When designing the parameter on the multiple quality characteristics, there has been a study for optimization of problems, but there has been few former study on the possible conflicting phenomena in considertion of the correlations among the characteristics. To solve the issue on the optimal design for multiple quality characteristics, the study propose the expected loss function with cross-product terms among the characteristics and derived range of the coefficients of terms. Therefore, the analysis have to be required a multivariate statistical technique. This paper introduces to multivariate capability indices and then selects a multivariate process capability index incorporated both the process variation and the process deviation from target among these indices under the multivariate normal distribution. We propose a new multivariate capability index MCpm using quality loss function instead of the process variation and this index is compared with the proposed indices when quality characteristics are independent and dependent of each other.

In the interest of deriving regressor that is robust to outliers, we propose a support vector regression (SVR) based on non-convex quadratic insensitive loss function with flexible coefficient and margin. The proposed loss function can be approximated by a difference of convex functions (DC). The resultant optimization is a DC program. We employ Newton’s method to solve it. The proposed model can explicitly enhance the robustness and sparseness of SVR. Numerical experiments on six benchmark data sets show that it yields promising results.

Power function distribution is a flexible lifetime distribution that may offer a good fit to some failure data sets. In this paper, we obtain Bayesian estimators of the shape parameter of Power function distribution. For the Posterior distribution of this parameter, we consider Exponential Prior, Pareto Prior, Chi-Square Prior, Quasi Prior and Extension of Jeffrey`s Prior. The three loss functions taken up are Squared Error Loss Function (SELF), Quadratic Loss Function (QLF) and Precautionary Loss Function (PLF). The performance of an estimator is assessed on the basis of its relative Posterior risk. Monte Carlo Simulations are used to compare the performance of the estimators. It is discovered that the PLF produces the least Posterior risk when Exponential and Pareto Priors are used. SELF is the best when Chi-Square, Quasi and Extension of Jeffrey`s Priors are used.

This study was done to establish clothing sizing system which can provide good fitting of apparel and minimize the loss due to excess inventory. One demensional mesurments(R. Martin) were made with the subjects, who were girls aged 11 to 12 living in Pusan and Kyungnam. Further the above-said mesurements were used to determine optimal clothing sizes using loss function for establishment of clothing sizing system. According to the results, optimal sizes of upper clothing for the late elementary schoolgirls should include 7 sizes, which will include heightbust circumferencesleeve length. And optimal sizes of lower clothing for the late elementary schoolgirls should include 7 sizes, which will include heightbust circumferencehip circumference.

In plants, glycoproteins have been implicated in a wide variety of cellular processes including production of cell walls, pollination, pathogen defence and cell-to-cell communication. In addition, they have attracted considerable interest from the medical field as the main cause of food and pollen allergies. N-glycosylation is a complex process that encompasses the biosynthesis and modification of sugar moieties in the endoplasmic reticulum (ER) and Golgi. The core oligosaccharide Glc3Man9GlcNAc2 is assembled by a series of membrane-bound glycosyltransferases as the lipid carrier dolichylpyrophosphate-linked glycan in the endoplasmic reticulum (ER). The first step of this assembly pathway on the ER luminal side is mediated by ALG3 (asparagine-linked glycosylation 3), which is a highly conserved reaction among eukaryotic cells. In Arabidopsis ALG3 mutant (alg3), an immature lipid-linked oligosaccharide structure, M5ER, was synthesized and efficiently processed into complex-type glycans. Although no high-mannose-type glycoproteins are detected in alg3 plants, these plants do not show a growth phenotype under normal growth conditions. However, the glycosylation abnormalities result in activation of marker genes diagnostic of the unfolded protein response and alg3 mutant showed a stress sensitive phenotype. These results indicate that ALG3 is a critical factor for correct N-glycosylation of proteins and is involved in the ER stress response.

데이터 불균형은 분류 문제에서 흔히 마주치는 문제로, 데이터셋 내의 클래스간 샘플 수의 현저한 차이에서 기인 한다. 이러한 데이터 불균형은 일반적으로 분류 모델에서 과적합, 과소적합, 성능 지표의 오해 등의 문제를 야기한다. 이를 해결하기 위한 방법으로는 Resampling, Augmentation, 규제 기법, 손실 함수 조정 등이 있다. 본 논문에서는 손실 함수 조정에 대해 다루며 특히, 불균형 문제를 가진 Multi-Class 블랙박스 동영상 데이터에서 여러 구성의 손실 함수(Cross Entropy, Balanced Cross Entropy, 두 가지 Focal Loss 설정: α = 1 및 α = Balanced, Asymmetric Loss)의 성능을 I3D, R3D_18 모델을 활용하여 비교하였다.

Data imbalance is a common issue encountered in classification problems, stemming from a significant disparity in the number of samples between classes within the dataset. Such data imbalance typically leads to problems in classification models, including overfitting, underfitting, and misinterpretation of performance metrics. Methods to address this issue include resampling, augmentation, regularization techniques, and adjustment of loss functions. In this paper, we focus on loss function adjustment, particularly comparing the performance of various configurations of loss functions (Cross Entropy, Balanced Cross Entropy, two settings of Focal Loss: α = 1 and α = Balanced, Asymmetric Loss) on Multi-Class black-box video data with imbalance issues. The comparison is conducted using the I3D, and R3D_18 models.

To select the best Third-Party Logistics(TPL) transportation suppliers(companies) it is essential to make a trade off between quantitative and qualitative factors. Considering different factors selection of the appropriate companies becomes more complicated. In the study, a Taguchi function is applied to each factor with which a loss can be quantified. For each candidate company the overall Taguchi loss is obtained based upon which the priority for selection is easily determined. Also, we present a mathematical model to calculate the optimal amount of goods to be transported from each of the company selected by the model. The Taguchi method and the model constructed will be used as a practical tool for the efficient decision making.

본 연구에서는 Vision Transformer 기반의 Anomaly Detection and Localization (VT-ADL) 모델에 초점을 맞추고, 손실 함수의 변경이 MVTec 데이터셋에 대한 이상 검출 및 지역화 성능에 미치는 영향을 비교 분석한다. 기존의 손실 함수를 KL Divergence와 Log-Likelihood Loss의 조합인 VAE Loss로 대체하여, 성능 변화를 심층적으로 조사했다. 실험을 통해 VAE Loss로의 전환은 VT-ADL 모델의 이상 검출 능력을 현저히 향상시키며, 특히 PRO-score에서 기존 대비 약 5%의 개선을 보였다는 점을 확인하였다. 이러한 결과는 손실 함수의 최적화가 VT-ADL 모델의 전반적인 성능에 중요한 영향을 미칠 수 있음을 시사한다. 또한, 이 연구는 Vision Transformer 기반 모델의 이상 검출과 지역화 작업에 있어서 손실 함수 선택의 중요성을 강조하며, 향후 관련 연구에 유용한 기준을 제공할 수 있을 것으로 기대된다.