• Journal of Broadcast Engineering
• /
• v.25 no.6
• /
• pp.980-993
• /
• 2020

Deep-learning based style transfer has recently attracted great attention, because it provides high quality transfer results by appropriately reflecting the high level structural characteristics of images. This paper deals with the problem of providing more stable and more diverse style transfer results of such deep-learning based style transfer method. Based on the investigation of the experimental results from the wide range of hyper-parameter settings, this paper defines the problem of the stability and the diversity of the style transfer, and proposes a partial loss normalization method to solve the problem. The style transfer using the proposed normalization method not only gives the stability on the control of the degree of style reflection, regardless of the input image characteristics, but also presents the diversity of style transfer results, unlike the existing method, at controlling the weight of the partial style loss, and provides the stability on the difference in resolution of the input image.

• Phonetics and Speech Sciences
• /
• v.12 no.4
• /
• pp.63-71
• /
• 2020

Speech recognition is one of the areas actively commercialized using deep learning and machine learning techniques. However, the majority of speech recognition systems on the market are developed on data with limited diversity of speakers and tend to perform well on typical adult speakers only. This is because most of the speech recognition models are generally learned using a speech database obtained from adult males and females. This tends to cause problems in recognizing the speech of the elderly, children and people with dialects well. To solve these problems, it may be necessary to retain big database or to collect a data for applying a speaker adaptation. However, this paper proposes that a new end-to-end speech recognition method consists of an acoustic augmented recurrent encoder and a transformer decoder with linguistic prediction. The proposed method can bring about the reliable performance of acoustic and language models in limited data conditions. The proposed method was evaluated to recognize Korean elderly and children speech with limited amount of training data and showed the better performance compared of a conventional method.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.17 no.5
• /
• pp.825-832
• /
• 2022

In this paper, we discuss the design and implementation of predictive and diagnostic models for realizing intelligent predictive models by collecting and storing the power output of agricultural photovoltaic power generation systems. Our model predicts the amount of photovoltaic power generation using RNN, LSTM, and GRU models, which are recurrent neural network techniques specialized for time series data, and compares and analyzes each model with different hyperparameters, and evaluates the performance. As a result, the MSE and RMSE indicators of all three models were very close to 0, and the R2 indicator showed performance close to 1. Through this, it can be seen that the proposed prediction model is a suitable model for predicting the amount of photovoltaic power generation, and using this prediction, it was shown that it can be utilized as an intelligent and efficient O&M function in an agricultural photovoltaic system.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2022.06a
• /
• pp.1265-1267
• /
• 2022

방대한 분량의 콘텐츠 홍수 속에서 원하는 소재를 찾기 위해 콘텐츠 내용을 검색할 수 있는 효과적인 방법이 지원되는 것은 창작을 자유롭게 하고, 콘텐츠 활용도를 높이기 위해 매우 중요하다. KBS 바다 서비스의 경우 분류체계 방법을 사용하고 있으나. 최근 딥러닝을 이용한 인공지능 기술의 발전으로 콘텐츠의 내용을 인공지능 기술로 태깅하고, 태깅된 텍스트 정보를 이용하여 검색할 수 있는 기술 개발이 활발히 수행되고, 국가적으로도 해당 기술을 지원하고 있다. 본 논문에서는 이러한 기술 개발의 선행 요소인 방송사의 제작과정에서 요구되는 동영상 소재 콘텐츠 검색의 요구사항을 KBS 비디오 아카이브 검색 키워드 실제 사용 데이터를 이용하여 분석하였다. 약 1,000여건의 검색 키워드 분석과 이용자와 운영자의 응답 내용을 고찰한 결과, 특정 키워드에 집중하여 검색할 수 있도록 보완하여 주는 것이 필요함을 알아내었다. 또한, 검색 범위를 효과적으로 축소하여 검색을 손쉽고 빠르게 할 수 있는 방법을 고찰하였다. 본 논문에서는 미디어 창작에서 필요한 소재 콘텐츠를 찾기 위해 연구 개발해야 할 미디어 속성 추출 기술의 방향성을 제시하였다.

• Journal of the Microelectronics and Packaging Society
• /
• v.30 no.3
• /
• pp.11-19
• /
• 2023

Recently with the rapid advancement of artificial intelligence (AI) technologies such as Chat GPT, AI semiconductors have become important. AI technologies require the ability to process large volumes of data quickly, as they perform tasks such as big data processing, deep learning, and algorithms. However, AI semiconductors encounter challenges with excessive power consumption and data bottlenecks during the processing of large-scale data. Thus, the latest packaging technologies are required for AI semiconductor computations. In this study, the authors have described packaging technologies applicable to AI semiconductors, including interposers, Through-Silicon-Via (TSV), bumping, Chiplet, and hybrid bonding. These technologies are expected to contribute to enhance the power efficiency and processing speed of AI semiconductors.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.18 no.6
• /
• pp.277-283
• /
• 2018

Some innovative researchers have had a dream of self-thinking intelligent computer. Alphago, at last, showed its possibility. With it, most computer engineers including even students can learn easily how to do it. As the interest to the deep learning has been growing, people's expectation is also naturally growing. In this research, we tried to enhance the game ability of a 7-poker system by applying machine learning techniques. In addition, we also tried to apply emotion analysis of a player to trace ones emotional changes. Methods and outcomes are to be explained in this paper.

• Journal of the Korean Society of Visualization
• /
• v.20 no.2
• /
• pp.21-27
• /
• 2022

Air purifiers can be placed where the air cleaning is required, making it easy to manage indoor air quality. The position of the air purifier affects the indoor airflow pattern, resulting in different air cleaning efficiency. Many efforts and strategies have been examined through numerical simulations and experiments to find the proper location of the air purifier, but problems still remain due to the various geometrical indoor spaces and arrangements. Herein, we develop an artificial intelligence model to predict the performance of an air purifier depending on the installation location. To obtain the training data, numerical simulations were performed on the different locations of the air purifiers and airflow patterns. The trained artificial intelligence model predicted the air exchange performance depending on the installation location of the air purifier with a prediction accuracy of 92%.

• Annual Conference on Human and Language Technology
• /
• 2021.10a
• /
• pp.3-8
• /
• 2021

인공지능 기술이 발달함에 따라 기계번역 기술도 많은 진보를 이루었지만 여전히 기계번역을 통한 번역문 내에는 사람이 교정해야 하는 오류가 다수 포함되어있다. 이렇게 번역 모델에서 생성되는 오류를 교정하는 전문인력의 요구를 경감시키기 위하여 기계번역 사후교정 연구가 등장하였고, 해당 연구는 현재 WMT를 주축으로 활발하게 연구되고 있다. 이러한 사후교정 연구는 최근 도메인 특화 관점에서 주로 연구가 이루어지고 있으며 현재 많은 도메인에서 유의미한 성과를 내고 있다. 하지만 이런 연구들은 기존 번역문의 품질을 얼만큼 향상시켰는가에 초점을 맞출 뿐, 다른 도메인 특화 번역모델의 성능과 비교했을 때 얼마나 뛰어난지는 밝히지 않기 때문에 사후교정 연구가 도메인 특화에서 효과적으로 작용하는지 명확하게 알 수 없다. 이에 본 연구에서는 도메인 특화 번역 모델과 도메인 특화 사후교정 모델간의 성능을 비교함으로써, 도메인 특화에서 사후교정을 통해 얻을 수 있는 실제적인 성능을 검증한다. 이를 통해 사후교정이 도메인 특화 번역모델과 비교했을 때 미미한 수준의 성능을 보임을 실험적으로 확인하였고, 해당 실험 결과를 분석함으로써 향후 도메인특화 사후교정 연구의 방향을 제안하였다.

• Asia-pacific Journal of Multimedia Services Convergent with Art, Humanities, and Sociology
• /
• v.9 no.8
• /
• pp.783-790
• /
• 2019

Cognitive science attempts to explain human intelligence on the basis of success of artificial neural network, which is called connectionism. The neural network, e.g., deep learning, seemingly promises connectionism to go beyond what it is. But those(Fodor & Pylyshyn, Fodor, & McLaughlin) who advocate classical computationalism, or symbolism claim that connectionism must fail since it cannot represent the relation between human thoughts and human language. The neural network lacks systematicity, so any output of neural network is at best association or accidental combination of data plugged in input units. In this paper, I first introduce structure of artificial neural network and what connectionism amounts to. Second, I shed light on the problem of systematicity the classical computationalists pose for the connectionists. Third, I briefly introduce how those who advocate connectionism respond to the criticism while noticing Smolensky's theory of vector product. Finally, I examine the debate of computationalism and connectionism on systematicity, and show how the problem of systematicity contributes to the development of connectionism and computationalism both.

• KIPS Transactions on Computer and Communication Systems
• /
• v.9 no.12
• /
• pp.321-332
• /
• 2020

Decentralized approaches are extensively researched by academia and industry in order to cover up the flaws of existing systems in terms of data privacy. Blockchain and decentralized learning are prominent representatives of a deconcentrated approach. Blockchain is secure by design since the data record is irrevocable, tamper-resistant, consensus-based decision making, and inexpensive of overall transactions. On the other hand, decentralized learning empowers a number of devices collectively in improving a deep learning model without exposing the dataset publicly. To motivate participants to use their resources in building models, a decent and proportional incentive system is a necessity. A centralized incentive mechanism is likely inconvenient to be adopted in decentralized learning since it relies on the middleman that still suffers from bottleneck issues. Therefore, we design an incentive model for decentralized learning applications by leveraging the Ethereum smart contract. The simulation results satisfy the design goals. We also outline the concerns in implementing the presented scheme for sensitive data regarding privacy and data leakage.