Deep Structured Learning: Architectures and Applications

이수욱 국제문화기술진흥원 2018 International Journal of Advanced Culture Technolo Vol.6 No.4

Deep learning, a sub-field of machine learning changing the prospects of artificial intelligence (AI) because of its recent advancements and application in various field. Deep learning deals with algorithms inspired by the structure and function of the brain called artificial neural networks. This works reviews basic architecture and recent advancement of deep structured learning. It also describes contemporary applications of deep structured learning and its advantages over the treditional learning in artificial interlligence. This study is useful for the general readers and students who are in the early stage of deep learning studies.

Identifying depression in the National Health and Nutrition Examination Survey data using a deep learning algorithm

Oh, Jihoon,Yun, Kyongsik,Maoz, Uri,Kim, Tae-Suk,Chae, Jeong-Ho Elsevier 2019 Journal of affective disorders Vol.257 No.-

<P><B>Abstract</B></P> <P><B>Background</B></P> <P>As depression is the leading cause of disability worldwide, large-scale surveys have been conducted to establish the occurrence and risk factors of depression. However, accurately estimating epidemiological factors leading up to depression has remained challenging. Deep-learning algorithms can be applied to assess the factors leading up to prevalence and clinical manifestations of depression.</P> <P><B>Methods</B></P> <P>Customized deep-neural-network and machine-learning classifiers were assessed using survey data from 19,725 participants from the NHANES database (from 1999 through 2014) and 4949 from the South Korea NHANES (K-NHANES) database in 2014.</P> <P><B>Results</B></P> <P>A deep-learning algorithm showed area under the receiver operating characteristic curve (AUCs) of 0.91 and 0.89 for detecting depression in NHANES and K-NHANES, respectively. The deep-learning algorithm trained with serial datasets (NHANES, from 1999 to 2012), predicted the prevalence of depression in the following two years of data (NHANES, 2013 and 2014) with an AUC of 0.92. Machine learning classifiers trained with NHANES could further predict depression in K-NHANES. There, logistic regression had the highest performance (AUC, 0.77) followed by deep learning algorithm (AUC, 0.74).</P> <P><B>Conclusions</B></P> <P>Deep neural-networks managed to identify depression well from other health and demographic factors in both the NHANES and K-NHANES datasets. The deep-learning algorithm was also able to predict depression relatively well on new data set—cross temporally and cross nationally. Further research can delineate the clinical implications of machine learning and deep learning in detecting disease prevalence and progress as well as other risk factors for depression and other mental illnesses.</P> <P><B>Highlights</B></P> <P> <UL> <LI> Estimating epidemiological contributors to depression and predicting the prevalence of depression are still challenging. </LI> <LI> We aimed to estimate factors affecting depression in National Health and Nutrition Examination Survey (NHANES) datasets using deep learning and machine learning algorithms. </LI> <LI> Deep-learning achieved a high performance for identifying depression on the NHANES datasets of both the United States and South Korea. </LI> <LI> Trained deep-learning and machine learning algorithms are useful for estimating the prevalence of depression. </LI> </UL> </P>

딥러닝 관련 발명의 특허법상 보호 방안에 대한 연구

전용철(Jun, Yong-Cheul) 동아대학교 법학연구소 2020 東亞法學 Vol.- No.86

딥러닝 기술은 컴퓨터가 사물 등을 구분하기 위한 개념을 가지기 위해 ‘학습’을 수행하며, ‘학습’을 수행하기 위한 학습 모델이 가지는 구조가 깊은 것을 특징으로 한다. 딥러닝 기술은 인간이 개념을 받아들이기 위한 신경적 구조와 활동을 컴퓨터 기술에 접목한 신경망 기술을 기반으로 한다. 최근 그래픽 카드 등 하드웨어의 연산처리 능력이 높아지고 신경망 기술의 이론적 뒷받침이 이루어지면서 딥러닝 기술의 성과가 두드러지게 나타나게 되었다. 본 논문에서는 딥러닝 기술이 가지는 특성을 파악하기 위해 딥러닝 기술에 관한 변천과 주요 딥러닝 기술에 대한 내용 및 이와 관련된 특허권들의 청구범위를 살펴보았다. 또한, 본 논문에서는 딥러닝 기술이 가지는 특성을 감안하여 딥러닝 모델 구조가 변경된 경우와 기존 딥러닝 모델이 특유 목적에 맞게 접목된 경우로 구분하여 딥러닝 관련 발명의 특허 등록 가능성 확보 방안을 검토하였으며, 특허 등록 후 침해 주장 시 입증 용이성을 확보하기 위한 실무적 방안에 대해 검토하였다. 딥러닝 기술과 관련한 발명의 경우 발명의 성립성과 진보성 요건 등 특허성 요건을 판단함에 있어서 컴퓨터 관련 발명의 일종으로 취급될 수 있다. 최근 컴퓨터 관련 발명의 특허 적격성 판단과 관련하여 미국에서의 Alice 판결 등 주요 판례가 주목받고 있고 이에 따라 미국과 우리나라에서의 심사기준에도 변동이 있어 본 논문에서는 컴퓨터 관련 발명의 특허 적격성 판단에 대한 미국 판례의 변천을 살펴본 후 우리나라의 특허법 규정 및 2019년 3월 개정된 특허청의 특허․실용신안 심사기준을 검토하였다. Deep learning is a technology to perform “learning” so that computers can adopt concepts to distinguish objects and the like. Deep-learning technology is characterized in that the structures of learning models for performing “learning” are “deep”. Deep-learning technology is based on neural-network technology in which neural structures and activities that enable human beings to accommodate concepts are grafted onto the computer technology. Recently, neural network technology has been underpinned by theory owning to the increased operational processing capacity of hardware units such as graphics cards, and remarkable performance of deep-learning technology has been exhibited. In order to understand the properties of deep-learning technology, changes in deep-learning technology, major aspects of deep-learning technology, and the claims of patented inventions relating to deep-learning technology are reviewed in this paper. Also, considering the properties of deep-learning technology, cases are divided into cases where deep-learning model structures are changed and cases where conventional deep-learning models are grafted onto deep-learning technology so as to comply with specific objectives. On this basis, how to secure the patentability of deep-learning-related inventions is reviewed in this paper, and how to easily demonstrate patent infringement after the deep-learning-related inventions are patented is also reviewed in practical terms. In the case of deep-learning-related inventions, they may be handled as a kind of computer-related inventions when the patentability requirements thereof are determined in terms of whether the subject matters thereof establishes inventions and they meet inventiveness requirements, etc.. Recently, with respect to the patent eligibility determination of computer-related inventions, important precedents including the Alice Corp. judgment in the United States have drawn attention, and examination guidelines in the United States and Korea have changed accordingly. In this regard, in this paper I review changes in US precedents on patent eligibility determination of computer-related inventions, and then review the Korean Patent Act and patent and utility model examination guidelines, which were revised in March 2019 by the Korean Intellectual Property Office.

인공지능 딥러닝을 활용한 조류현상 예측기술 개발 및 활용방안

홍한움,조을생,강선아,한국진 한국환경정책평가연구원 2020 기본연구보고서 Vol.2020 No.-

Ⅰ. 연구의 배경 및 목적 1. 연구 개요 □ 연구명: 인공지능 딥러닝을 이용한 조류현상 예측기술 개발 및 활용방안 □ 연구기간: 2020.1.1~2020.12.31. 2. 연구의 필요성 및 목적 □ 조류현상의 원인 ㅇ 조류현상은 녹조현상과 적조현상을 포함 - 녹조현상: 강 및 호수에 남조류 과다 생성 - 적조현상: 바다에 갈색을 띠는 규조류 및 와편모조류 번성 □ 현행 조류경보제의 한계 ㅇ 환경부 및 국립환경과학원에서는 유해남조류 실측치 및 EFDC 모형에 기반하여 조류 경보제 시행 ㅇ 물리 모형의 한계 - 탄탄한 이론적 배경을 기반으로 하나, 모형이 요구하는 세밀한 데이터를 확보하는데 한계가 있음 - 질량보존의 법칙에 기반한 물리 모형 활용 생명활동인 조류현상 예측에 한계가 있음 ㅇ 딥러닝 예측을 대안 및 보완방안으로 고려 Ⅱ. 현행 녹조대응정책 1. 조류경보제 □ 도입 시기: 1998년 □ 법적 근거: 「물환경보전법」 제21조 □ 대상 ㅇ 상수원 및 친수활동구간 28개소 지점 ㅇ 발령권자: 국립환경과학원 □ 분석 항목 ㅇ 유해남조류세포수 실측치 ㅇ 상수원 구간 기준 - 관심: 1,000(cells/mL) 이상 - 경계: 10,000(cells/mL) 이상 - 대발생: 1,000,000(cells/mL) 이상 2. (구) 수질예보제 □ 도입 시기: 2012년 □ 법적 근거: 「물환경보전법」 제21조 □ 대상 ㅇ 4대강 16개 보 및 북한강 삼봉리 등 17개 지점 ㅇ 발령권자: 국립환경과학원 □ 분석 항목 ㅇ 수온 및 클로로필a 농도 예측치 ㅇ 현재는 조류경보제와 수질예보제를 통합 운영함에 따라 예측은 수행하지만 예보 발령은 하지 않음 □ 현재 유해남조류 예측정보 제공 ㅇ 현재 국립환경과학원에서 유해남조류 예측정보 제공 중 ㅇ 주 2회 월·목요일 조류경보제 6개 지점 대상 ㅇ 유해남조류세포수 및 수온예측 결과 발표 3. 수질측정망 현황 □ 법적 근거 ㅇ 「환경정책 기본법」 제22조 및 「물환경보전법」 제9조 □ 구성 ㅇ 수질측정망 - 대상: 하천, 호소, 농엽용수, 도시관류, 산단하천에서의 수질측정데이터 - 제공 정보: 수심, 수소이온농도, 용존산소량, BOD, COD, 부유물질, 총질소, 총인, 총유기탄소(TOC), 수온, 페놀류, 전기전도도, 총대장균군수, 용존총질소, 암모니아성 질소, 질산성 질소, 용존총인, 용존총인, 인산염인, 클로로필a, 투명도 - 주기: 월 1회, 주요지점에 대해서는 주 1회 ㅇ 총량측정망 - 대상: 수질오염총량제 대상 지역에 대해 총량 관리에 필요한 기초데이터 - 제공 정보: 수온, 수소이온농도, 전기전도도, 용존산소, BOD, COD, 부유물질, 총질소, 총인, TOC, 유량 - 주기: 월 1회 ㅇ 자동측정망 - 수동적으로 측정되는 일반측정망의 보완을 위해 운영 - 제공 정보: (공통항목) 수온, 수소이온농도, 용존산소량, 전기전도도, TOC (선택항목) 탁도, 클로로필a, TN, TP, NH3-N, NO3-N, PO4-P, VOCs(9종 10개), 페놀, 중금속, 생물감시항목 - 주기: 일 1회 ㅇ 퇴적물측정망 - 목적: 국 하천 및 호소 등 수질보전대상 공공수역에 대한 퇴적물의 이화학적 특성 조사 - 제공 정보: (공통항목) 수온, 수소이온농도, 용존산소량, 전기전도도, TOC (선택항목) 채취시간 최고수심, 표층측정수심, 표층 및 저층 수심, 수온, 용존산소량, pH, 전기전도도, 퇴적물 입도, 함수율, 완전연소가능량 비율 및 등급, COD, TOC, TN, TN등급, TP, SRP, 중금속, 보존성 원소 농도 - 주기: (하천) 상·하반기 연 2회, (호소) 연 1회 ㅇ 이 외에 방사성 측정망 및 생물측정망 추가 측정 Ⅲ. 수질 예측 모형 1. 물리모형 □ 모형 예시 ㅇ EFDC, QUAL2K, WASP 등 ㅇ 국립환경과학원에서는 EFDC 기반 모형 운용 중 □ 구성 ㅇ 수계를 소구역으로 분할한 모델 격자망 구성 및 경계조건 설정 ㅇ 격자망 안의 소구역 단위에서 수질 추정 □ 사례 ㅇ 신창민 외(2017)의 EFDC 활용 영산강 수계 예측 2. 딥러닝 모형 □ 모형 구조 ㅇ 다층 퍼셉트론(DMLP) - 신경망의 뉴런과 시냅스를 모방한 모형. 입력층, 은닉층, 출력층으로 구성. 은닉층을 여러 개 두는 다층 구조로 구성 ㅇ 순환신경망(RNN) - 다층 퍼셉트론 모형에서 전 시점 은닉 노드의 피드백을 추가 반영한 모형 - 현대에는 단순 순환신경망 모형을 기반으로 하여 장기 기억을 활용할 수 있는 GRU, LSTM 모형을 활용 □ 수질 예측 모형 적용 예시 3. 물리모형 vs 딥러닝 알고리듬 □ 물리모형 ㅇ 잘 확립된 수학/물리법칙에 기반 ㅇ 실제 관측값은 모형 평가에 활용 ㅇ 물리적 방정식을 통해 관측값보다 세밀한 해상도에서 예측수행 가능 ㅇ 단점 - 불확실한 초기/경계조건으로 인한 오차 - 이상현상을 예측하기 어려움 - 부실한 입력데이터, 모델 관계식의 불안정성, 모델링 방법 등의 문제로 작동하지 않을 수 있음 □ 딥러닝 알고리듬 ㅇ 기계학습을 통해 입력변수와 출력변수의 관계 구축 ㅇ 실제 관측값 모형 구축에 활용 ㅇ 측정값의 오차를 정량화하여 모형 안에 오류 조건 내장 ㅇ 물리 모형 대비 불확실성이 큰 단기 예측에 강점 ㅇ 단점 - 많은 데이터 요구 - 관측 해상도보다 세밀화 불가능 - 입력변수와 출력변수의 관계를 설명할 수 없기 때문에 실제 활용에 한계가 있음 Ⅳ. 딥러닝 기반 조류예측 알고리듬 개발 1. 데이터 수집 및 전처리 □ 모형 구축 대상 ㅇ 대상 지점: 한강 친수활동구간 조류 관찰지점 ㅇ 대상 변수 - 조류현상의 직접 원인인 유해남조류세포수 직접 예측 - 클로로필a 예측 등을 통해 우회적으로 녹조현상을 예측한 선행연구와 차별성이 있음 □ 모형 구축 기간 ㅇ 대상기간: 2007.4~2020.8. ㅇ 조류현상으로부터 비교적 안전한 겨울철인 12 ~ 3월의 겨울철 데이터는 관측값이 없으므로 제외 ㅇ 수집 데이터 2. 조류 데이터 특성 □ 기술통계 □ 특징 ㅇ 극단적으로 오른쪽으로 치우친 비대칭 분포 ㅇ 온도가 높은 여름철에 집중적으로 발생하여 이와 같은 극단적인 비대칭 특성이 나타남 ㅇ 극단적 비대칭 특성으로 인해 물리 모형이나 전통적인 통계 모형 등으로 유해남조류를 직접 예측하기 어려움 3. 예측 알고리듬 개발 □ RNN 모형 구축 ㅇ 로그스케일에서의 유해남조류세포수를 예측대상으로 함 ㅇ 장기기억 정보 활용을 위해 LSTM 예측 알고리듬을 구축함 ㅇ 최적화를 위한 손실함수: 최소제곱함수 최적화 알고리듬: ADAM ㅇ 학습 데이터(training data): 2007.4~2016.11. 검정 데이터(test data): 2017.4~2020.6. □ 결과 ㅇ 전 관측소에서 유해남조류의 증감 패턴을 잘 예측함. 친수활동구간은 하천 하류에 위치하여 데이터 불안정성이 커 전통적인 예측 방법으로는 예측하기 어려우나, 본 연구에서는 증감 패턴을 잘 예측함 ㅇ 가장 큰 극단값의 발생을 동 시점에서 예측함 ㅇ 예측오차 Ⅴ. 결론 및 학술적 성과 □ 학술적 성과 ㅇ 물리모형을 활용한 예측은 명확한 이론을 바탕으로 정립되어 있기 때문에 수온, 용존산소량, 총인, 총질소 등의 수질 예측에 널리 쓰임. 하지만 질량보존법칙을 기저로 하는 물리 방정식을 활용한 예측은 보존성 물질에는 잘 맞으나 살아 있는 생명체의 활동인 조류현상 예측에는 한계가 있음 ㅇ 기존 조류현상 예측 연구는 조류현상의 직접적인 원인인 유해남조류세포수(cells/mL)를 직접 예측하지 않고 클로로필a 농도(mg/m3) 예측 결과를 활용하는 것으로 대체함 ㅇ 본 연구에서는 물리 모형으로는 예측하기 힘든 유해남조류세포수 예측에 대한 대안으로 순환신경망 기반의 딥러닝 알고리듬을 활용함. 조류 증감 및 이상현상 발생을 동 시점에서 잘 예측함 □ 한계 ㅇ 입력변수로 수질, 상류 수질, 수위, 기상 정보만을 활용하였는데, 이는 물리 모형에서 이미 쓰고 있는 변수 위주임. 인구 변화와 같은 인간 사회 활동에 관한 변수를 고려하면 딥러닝 분석의 이점을 더 크게 활용할 수 있음. 위성 이미지 등의 비정형 데이터 또한 추가로 고려할 수 있음 ㅇ 데이터 수의 한계. 본 연구에서는 2007년부터 2016년까지의 총 365개 주별 데이터를 활용하여 모형을 학습하였는데, 이 데이터 수 자체가 충분하다고 할 수 없음. 새로운 데이터가 추가될 때마다 예측 모형을 업데이트하여 효율성을 높여야 함 ㅇ 딥러닝 모형의 근본적 한계. 실제 모형의 자세한 동작 과정을 명확히 알 수 없다는 블랙박스(black-box) 특성으로 인한 한계가 있음. 정책을 시행할 때는 근거가 필요한데, 딥러닝 예측 모형의 블랙박스 특성은 명확한 근거를 수립하기 어려움 □ 결론 및 제언 ㅇ 이미 구축된 모형에 대한 예측 수행은 매우 간단하므로 현재의 조류 예보에 참고 정보로 바로 활용할 수 있음 ㅇ 딥러닝 모형을 활용한 예측과 물리 모형을 활용한 예측 모두 장단점이 있으므로 두 예측 방식을 통합하는 것이 가장 바람직함. 딥러닝 모형에 기반을 두고 목적함수의 제약 조건에 물리 방정식을 포함하는 방식으로 물리 모형을 통합할 수도 있고, 물리모형 예측의 부분 모듈에 딥러닝 학습을 부분적으로 수행하는 방식으로 물리 모형을 기반으로 딥러닝 모형을 통합할 수도 있음 Ⅰ. Background and Aims of Research 1. Research outline □ Research title: Development and application of an algal bloom forecast system using artificial intelligence deep learning technology □ Research period: January 1, 2020 ~ December 31, 2020 2. Necessity and purpose of research □ Limitations of the current algal bloom warning system ㅇ The Ministry of Environment and the National Institute of Environmental Research implemented an algal bloom warning system based on the measured values of harmful blue-green algae and the EFDC model. ㅇ Limitations of physics-based models - They have a solid theoretical background but there is a difficulty in securing the detailed data required by the model. - Since algal blooms are living organisms, the law of conservation of mass does not apply to the number of harmful blue-green algae cells. Therefore, the physics-based model has limitations. - Deep learning-based forecasting can be considered as an alternative and a complementary method. Ⅱ. Current Algal Bloom Response Policy 1. Algal bloom warning system □ Year of introduction: 1998 □ Legal basis: Article 21 of the Water Environment Conservation Act □ Target ㅇ 28 branches of water supply sources and hydrophilic activities ㅇ Issuer: Basin Environmental Office and local governments □ Analysis items ㅇ Measured numbers of harmful blue-green algae cells ㅇ Based on water source section - Attention: 1,000 (cells/mL) or more - Alert: 10,000 (cells/mL) or more - Large bloom: 1,000,000 (cells/mL) or more ㅇ Based on hydrophilic activities section - Attention: 20,000 (cells/mL) or more - Alert: 100,000 (cells/mL) or more 2. (Former) Water quality forecast system □ Year of Introduction: 2012 □ Legal basis: Article 21 of the Water Environment Conservation Act □ Target ㅇ 17 branches including 16 barrages and the Bukhan River Sambong-ri of the four major rivers of South Korea ㅇ Issuer: National Institute of Environmental Research □ Analysis items ㅇ Predicted water temperature and chlorophyll-a concentration ㅇ Currently, as the algal bloom warning system and the water quality forecast system are integrated, no forecast is issued although forecasting is performed. □ Providing forecasts for harmful blue-green algae cells ㅇ Twice a week, Monday and Thursday, six branches that are targets of the algal bloom system ㅇ Issuing the predicted number of harmful blue-green algae cells and water temperature predictions 3. Status of the water quality monitoring network □ Legal basis ㅇ Article 22 of the Basic Act on Environmental Policy and Article 9 of the Water Environment Conservation Act □ Organization ㅇ Water quality monitoring network - Target: water quality measurement data in rivers, lakes, agricultural water, urban streams, and industrial rivers - Provided information: water depth, hydrogen ion concentration, dissolved oxygen content, BOD, COD, suspended matter, total nitrogen, total phosphorus, total organic carbon (TOC), water temperature, phenols, electrical conductivity, total coliform group, dissolved total nitrogen, ammonia nitrogen, nitrate nitrogen, dissolved total phosphorus, phosphate phosphorus, chlorophyll a, transparency - Cycle: once a month, once a week for major locations ㅇ Total quantity measurement network - Target: basic data for total amount management in areas subject to the total water pollution rate system - Provided information: water temperature, hydrogen ion concentration, electrical conductivity, dissolved oxygen, BOD, COD, suspended matter, total nitrogen, total phosphorus, TOC, flow rate - Cycle: once a month ㅇ Automatic measurement network - Operated to complement the hand-operated measurements of the water quality monitoring network - Provided information: (Common) water temperature, hydrogen ion concentration, dissolved oxygen content, electrical conductivity, TOC (Optional) Turbidity, chlorophyll a, TN, TP, NH₃-N, NO₃-N, PO₃-P, VOCs (nine types, ten items), phenol, heavy metals, biological monitoring items - Cycle: once a day ㅇ Sediment monitoring network - Purpose: investigation of the physicochemical properties of sediments in public waters subject to water quality conservation of South Korea - Provided information: (Common) water temperature, hydrogen ion concentration, dissolved oxygen content, electrical conductivity, TOC (Optional) maximum depth during collection, surface measurement depth, surface and bottom depth, water temperature, dissolved oxygen content, pH, electrical conductivity, sediment particle size, moisture content, ratio and grade of complete combustion potential, COD, TOC, TN, TN grade, TP, SRP, heavy metals, conservative element concentration - Cycle: (River) twice a year for the first and second halves, (Lake) once a year ㅇ In addition, there are additional measurements of radioactive monitoring networks and biometric networks. Ⅲ. Water Quality Prediction Models 1. Physics-based model □ Example ㅇ EFDC, QUAL2K, WASP, etc. ㅇ The National Institute of Environmental Research is operating an EFDC-based model. □ Organization ㅇ Construct a grid network by dividing the water system into sub-regions and set boundary conditions ㅇ Estimate the water quality in sub-area units within the grid 2. Deep learning algorithm □ Model structure ㅇ Multi-layer perceptron (MLP) - It mimics the neurons and synapses of a neural network. It consists of an input layer, a hidden layer, and an output layer. it has a multi-layered structure with more than one hidden layer. ㅇ Recurrent Neural Network (RNN) - It additionally reflects the feedback effects of previous hidden nodes. - Nowadays, GRU and LSTM models are used. These models utilize the long-term memory based on a simple recurrent neural network. 3. Physics-based model vs. Deep learning algorithm □ Physics-based model ㅇ Based on well-established mathematical/physical laws ㅇ Actual observations are used for model evaluation. ㅇ Prediction can be performed at a more detailed resolution than observed values based on physical equations. ㅇ Disadvantages - Errors due to uncertain initial/boundary conditions - Difficulty in predicting the abnormal phenomena - May not work due to problems such as poor input data, instability of model relations, modeling method, etc. □ Deep learning algorithm ㅇ Establish the relationship between input and output variables through machine learning ㅇ Actual observations are used for model construction. ㅇ Includes error conditions in the model by quantifying the error of the measurements ㅇ Advantages in short-term predictions with greater uncertainties compared to physics-based models ㅇ Disadvantages - Requires a huge amount of data - Cannot be performed at a more detailed resolution than observation resolution - Practical application is limited since the relationship between input and output variables cannot be explained. Ⅳ. Development of an Algal Bloom Forecast Algorithm Based on Deep Learning 1. Data collection and preprocessing □ Model construction target ㅇ Target point: algae observation point in the hydrophilic activity section of the Han River ㅇ Target variable - Direct prediction of the number of harmful blue-green algae cells which is the direct cause of the algal bloom - Differentiated from previous studies that indirectly predicted the algal bloom through chlorophyll a prediction □ Model construction period ㅇ Target period: April 2007 ~ August 2020 ㅇ Data in winter from December to March, which is relatively safe from algal blooms, are excluded. 2. Characteristics of algae data □ Descriptive statistics □ Characteristics ㅇ Extremely right-skewed asymmetric distribution ㅇ Extreme asymmetric distribution is exhibited since algal blooms occur intensively in summer when the temperature is high. ㅇ Because of this, it is difficult to directly predict harmful blue-green algae using physics-based models or traditional statistical models. 3. Development of a predicting algorithm □ RNN model construction ㅇ Target of prediction: the number of harmful blue-green algae cells ㅇ Constructing an LSTM prediction algorithm to utilize the long-term memory information ㅇ Loss function for optimization: least squares function Optimization algorithm: ADAM ㅇ Training data: April 2007 ~ November 2016 Test data: April 2017 ~ June 2020 □ Results ㅇ The increasing and decreasing patterns are well predicted although there is difficulty in predicting using traditional prediction methods due to high data instability, which results from the fact that the hydrophilic activity section is located downstream of the river. ㅇ Well predict the occurrence of the largest extreme value at the same time ㅇ Prediction error Ⅴ. Conclusion and Achievements □ Achievements ㅇ Since the prediction using a physical model is established based on a well-established theory, it is widely used to predict properties of water quality such as water temperature, dissolved oxygen, total phosphorus, and total nitrogen. The prediction using the physical equation based on the law of conservation of mass is well suited for conservative substance. However, there is a limitation in the prediction of algae cells since it is the activity of living organisms. ㅇ Existing algal phenomena prediction studies have not directly predicted the number of harmful blue-green algae cells, which is the direct cause of algal phenomena. It is replaced by using the results of chlorophyll a concentration prediction. ㅇ In this study, a deep learning algorithm based on recurrent neural networks was used as an alternative method to predict the number of harmful blue-green algae cells. It well predicted the increasing or decreasing patterns of algae and the occurrence of abnormal phenomena at the concurrent point. □ Limitations ㅇ Only water quality, upstream water quality, water level, and meteorological information were used as input variables. These variables are already used in the physical model. Taking into account social variables such as population change and the benefits of deep learning analytics can be leveraged to a greater extent. Unstructured information such as satellite images can be additionally considered. ㅇ There is a limitation in the amount of data. In this study, the model was studied using data from a total of 365 weekly data collections from 2007 to 2016, but this amount itself is not sufficient. Whenever new data are added, the predictive model should be updated to increase the prediction efficiency. ㅇ There is a limitation due to the black-box characteristic. The detailed operational process of the prediction model cannot be clearly observed. When implementing a policy, evidence is needed. The black-box characteristic of deep learning prediction models makes it difficult to provide clear evidence. □ Conclusions and suggestions ㅇ Because it is very simple to perform predictions with the model that has already been established, it can be directly used as reference information for current algal bloom forecasts. ㅇ Since predictions using deep learning models and physics-based models both have advantages and disadvantages, it is most desirable to integrate the two prediction methods. Based on the deep learning model, the physical model can be integrated by including the physical equation in the constraint of the objective function. Or, deep learning can be partially performed in the partial module of the physical model prediction.

딥러닝 예측 결과 정보를 적용하는 복합 미생물배양기를 위한 딥러닝 구조 개발

김홍직,이원복,이승호 한국전기전자학회 2023 전기전자학회논문지 Vol.27 No.1

In this paper, we develop a deep learning structure for a complex microbial incubator that applies deep learning prediction result information. The proposed complex microbial incubator consists of pre-processing of complex microbial data, conversion of complex microbial data structure, design of deep learning network, learning of the designed deep learning network, and GUI development applied to the prototype. In the complex microbial data preprocessing, one-hot encoding is performed on the amount of molasses, nutrients, plant extract, salt, etc. required for microbial culture, and the maximum-minimum normalization method for the pH concentration measured as a result of the culture and the number of microbial cells to preprocess the data. In the complex microbial data structure conversion, the preprocessed data is converted into a graph structure by connecting the water temperature and the number of microbial cells, and then expressed as an adjacency matrix and attribute information to be used as input data for a deep learning network. In deep learning network design, complex microbial data is learned by designing a graph convolutional network specialized for graph structures. The designed deep learning network uses a cosine loss function to proceed with learning in the direction of minimizing the error that occurs during learning. GUI development applied to the prototype shows the target pH concentration (3.8 or less) and the number of cells (10⁸ or more) of complex microorganisms in an order suitable for culturing according to the water temperature selected by the user. In order to evaluate the performance of the proposed microbial incubator, the results of experiments conducted by authorized testing institutes showed that the average pH was 3.7 and the number of cells of complex microorganisms was 1.7 × 10⁸. Therefore, the effectiveness of the deep learning structure for the complex microbial incubator applying the deep learning prediction result information proposed in this paper was proven.

Deep Structured Learning: Architectures and Applications

Soowook Lee 국제문화기술진흥원 2018 International Journal of Advanced Culture Technolo Vol.6 No.4

Deep learning, a sub-field of machine learning changing the prospects of artificial intelligence (AI) because of its recent advancements and application in various field. Deep learning deals with algorithms inspired by the structure and function of the brain called artificial neural networks. This works reviews basic architecture and recent advancement of deep structured learning. It also describes contemporary applications of deep structured learning and its advantages over the treditional learning in artificial interlligence. This study is useful for the general readers and students who are in the early stage of deep learning studies.

Deep Learning을 활용한 산사태 결정론 방법의 활용성 고찰

민대홍(Min, Dae-Hong),윤형구(Yoon, Hyung-Koo) 한국방재학회 2021 한국방재학회논문집 Vol.21 No.5

산사태 위험지역을 결정론적인 방법으로 도출할 수 있는 Analytic Hierarchy Process (AHP) 기반의 선행 연구가 2017년도에 제안되었다. 해당 연구의 목적은 기존에 제안된 결정론적인 방법의 활용성을 향상시키고자 deep learning 기법을 적용하여 해당 방법의 신뢰성을 검증하는 것이다. AHP 기반의 결정론적인 방법은 8개 인자인 세립분 함량, 표토층 두께, 간극비, 탄성계수, 전단강도, 투수계수, 포화도 그리고 함수비로 구성되며 이를 통해 안전율을 도출할 수 있다. 대상 지역을 1 m 정사각형의 격자로 구성한 후 현장 및 실내 실험을 통해 8개의 인자를 도출하였다. 안전율은 Mohr-Coulomb의 파괴 이론을 통해 계산하여 deep learning의 출력 값으로 활용하였다. Deep learning 기법 적용 시 입력 값과 출력 값의 학습 능률을 향상시키기 위하여 경사하강법 중 Bayesian regularization을 적용하였으며, 학습 결과 실제 안전율과 deep learning 기법으로 예측된 안전율이 train과 test 단계 모두에서 우수한 신뢰성을 보여준다. 해당 연구에서 활용한 deep learning 기법이 산사태 위험지역 선정에 결정론적 방법으로 유용하게 이용될 것으로 사료된다. A method for estimating landslide susceptibility based on the analytic hierarchy process (AHP) was developed in 2017 as a deterministic method. The objective of this study is to verify the reliability of the proposed method by applying deep learning to improve the applicability of the method. The AHP-based deterministic method comprises eight factors: fines content, soil thickness, porosity, elastic modulus, shear strength, hydraulic conductivity, saturation, and water content. After dividing the testing area into 1 m square grids, eight factors were derived through field and laboratory experiments. The factor of safety was calculated based on the Mohr-Coulomb failure theory. Finally, the input and output values of deep learning were obtained. Bayesian regularization was applied among gradient descents to improve the learning efficiency when applying machine learning. The actual and predicted factors of safety were compared, and they showed excellent reliability in both the training and test phases. This study demonstrates that the AHP-based deterministic method with deep learning is valuable for determining landslide risk areas.

딥러닝 기반 체육활동에 참여한 알파세대 학습경험이 성취도와 메타인지에 미치는 영향

홍선옥,강현욱 여가문화학회 2023 여가학연구 Vol.21 No.3

This study aims to investigate the impact of learning experience on achievement and metacognition, as well as the structural relationship among the alpha generation who participated in deep learning-based physical activities. The purpose of this study was to analyse the structural relationship among learning experience, achievement, and metacognition in alpha generation participating deep learning-based physical activities. Based on deep learning, flying disc, ball throwing, and passing movements were selected as target physical activities. Two physical education teachers modeled each activity using Teachable Machine over three days. Afterwards, during regular physical education classes and after-school sessions, learners' performances of flying disc, ball throwing, and passing movements were analyzed and feedback was provided through Teachable Machine. 350 questionnaires were distributed. To prevent incomplete or incorrect answers, teachers paraphrased the questionnaires into easier words, and enough explanation and modeling were provided. Total 350 answers were considered to be effective samples. SPSS Window Version 28.0 was used to analyze statistical data. Frequency analysis and correlation analysis were conducted. To test the hypothesis, AMOS 28.0 was used for confirmatory factor analysis and structural equation model analysis. The results are as follows; First, the alpha generation’s learning experience in deep learning-based physical activities has a meaningful influence on the achievement they have. Second, the alpha generation’s achievement in deep learning-based physical activities has a meaningful effect on the metacognition they have. Third, the alpha generation’s learning experience in deep learning-based physical activities also affects meaningfully on their metacognition. Collectively, this study confirms deep learning-based physical activities can be beneficial for alpha generation, since they improve achievement, and metacognition of alpha generation. 본 연구는 딥러닝(Deep Learning: DL)을 기반으로 체육활동에 참여한 알파세대의 학습경험과 성취도 및 메타인지의 관계를 구조적으로 알아보고자 한다. 이 연구에서는 총 350개의 설문지를 배포하여 모든 데이터를 최종 사용가능한 표본으로 선정하였다. 딥러닝을 기반으로 한 체육활동으로는 플라잉디스크와 공 던지기 및 패스 동작을 선정하였으며, 체육 교사 2인이 3일에 걸쳐 각 활동별 예시 동작을 Teachable Machine에 학습시켰다. 이후 정규 체육시간 및 방과후수업 시간에 학습자들의 플라잉디스크와 공 던지기 및 패스 동작 수행을 Teachable Machine을 통해 분석하고 피드백을 제공했다. 통계적 분석은 SPSS Window Version 28.0을 이용하여 빈도분석과 상관관계분석을 수행하였다. 또한 AMOS 28.0을 이용하여 확인적 요인분석을 실시하여 연구의 가설을 확인하였다. 마지막으로 구조방정식의 모델을 적용하여 다음과 같은 결론을 얻었다. 첫째, 딥러닝을 기반으로 한 체육활동에 참여한 알파세대의 학습경험은 성취도에 유의한 영향을 미치는 것으로 나타났다. 둘째, 딥러닝을 기반으로 한 체육활동에 참여한 알파세대의 성취도는 메타인지에 유의한 영향을 미치는 것으로 나타났다. 구조적 결과는 학습경험이 많을수록 참가자가 더 높은 수준의 성취도와 메타인지를 가지는 것으로 알 수 있었다. 셋째, 딥러닝을 기반으로 한 체육활동에 참여한 알파세대의 학습경험은 메타인지에 유의한 영향을 미치는 것으로 나타났다. 알파세대에게 딥러닝 기술을 활용한 체육활동은 학습자의 성취도 및 메타인지에 도움을 주는 것을 확인하였다.

학습률 적용에 따른 흉부영상 폐렴 유무 분류 비교평가

김지율,예수영 한국방사선학회 2022 한국방사선학회 논문지 Vol.16 No.5

This study tried to suggest the most efficient learning rate for accurate and efficient automatic diagnosis of medical images for chest X-ray pneumonia images using deep learning. After setting the learning rates to 0.1, 0.01, 0.001, and 0.0001 in the Inception V3 deep learning model, respectively, deep learning modeling was performed three times. And the average accuracy and loss function value of verification modeling, and the metric of test modeling were set as performance evaluation indicators, and the performance was compared and evaluated with the average value of three times of the results obtained as a result of performing deep learning modeling. As a result of performance evaluation for deep learning verification modeling performance evaluation and test modeling metric, modeling with a learning rate of 0.001 showed the highest accuracy and excellent performance. For this reason, in this paper, it is recommended to apply a learning rate of 0.001 when classifying the presence or absence of pneumonia on chest X-ray images using a deep learning model. In addition, it was judged that when deep learning modeling through the application of the learning rate presented in this paper could play an auxiliary role in the classification of the presence or absence of pneumonia on chest X-ray images. In the future, if the study of classification for diagnosis and classification of pneumonia using deep learning continues, the contents of this thesis research can be used as basic data, and furthermore, it is expected that it will be helpful in selecting an efficient learning rate in classifying medical images using artificial intelligence. 본 연구는 딥러닝을 이용한 흉부 X선 폐렴 영상에 대하여 정확하고 효율적인 의료영상의 자동진단을 위해서 가장 효율적인 학습률을 제시하고자 하였다. Inception V3 딥러닝 모델에 학습률을 0.1, 0.01, 0.001, 0.0001로 각각 설정한 후 3회 딥러닝 모델링을 수행하였다. 그리고 검증 모델링의 평균 정확도 및 손실 함수 값, Test 모델링의 Metric을 성능평가 지표로 설정하여 딥러닝 모델링의 수행 결과로 획득한 결과값의 3회 평균값으로 성능을 비교 평가하였다. 딥러닝 검증 모델링 성능평가 및 Test 모델링 Metric에 대한 성능평가의 결과, 학습률 0.001을 적용한 모델링이 가장 높은 정확도와 우수한 성능을 나타내었다. 이러한 이유로 본 논문에서는 딥러닝 모델을 이용한 흉부 X선 영상에 대한 폐렴 유무 분류 시 학습률을 0.001로 적용할 것을 권고한다. 그리고 본 논문에서 제시하는 학습률의 적용을 통한 딥러닝 모델링 시 흉부 X선 영상에 대한 폐렴 유무 분류에 대한 인력의 보조적인 역할을 수행할 수 있을 거라고 판단하였다. 향후 딥러닝을 이용한 폐렴 유무 진단 분류 연구가 계속해서 진행될 시, 본 논문의 논문 연구 내용은 기초자료로 활용될 수 있다고 여겨지며 나아가 인공지능을 활용한 의료영상 분류에 있어 효율적인 학습률 선택에 도움이 될 것으로 기대된다.

A Review of Deep Learning Research

( Ruihui Mu ),( Xiaoqin Zeng ) 한국인터넷정보학회 2019 KSII Transactions on Internet and Information Syst Vol.13 No.4

With the advent of big data, deep learning technology has become an important research direction in the field of machine learning, which has been widely applied in the image processing, natural language processing, speech recognition and online advertising and so on. This paper introduces deep learning techniques from various aspects, including common models of deep learning and their optimization methods, commonly used open source frameworks, existing problems and future research directions. Firstly, we introduce the applications of deep learning; Secondly, we introduce several common models of deep learning and optimization methods; Thirdly, we describe several common frameworks and platforms of deep learning; Finally, we introduce the latest acceleration technology of deep learning and highlight the future work of deep learning.