SlideShare a Scribd company logo
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
Hello, DL!
Introductiontodeeplearning
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
Robin
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
a table of contents
목차
1 Deeplearning이란?
2 CNN
3 GAN
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
Part 1,
DeepLearning이란?
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
DeepLearning이란?
MachineLearning
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
DeepLearning이란?
MachineLearning Neuron
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
DeepLearning이란?
MachineLearning Neuron DeepLearning
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
DeepLearning이란?
출처: https://blog.naver.com/PostView.nhn?blogId=dsjang650628&logNo=221864626337
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
DeepLearning의종류
Network schemes 에 따른 분류
CNN
(ConvolutionalNeuralNetwork)
GAN
(GenerativeAdversarialNetwork)
RNN
(RecurrentNeuralNetwork)
LSTM
(LongShort-TermMemorymodels)
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
DeepLearning의종류
Applications 에 따른 분류
CV
(ComputerVision)
NLP
(NaturalLanguageProcessing)
RF
(Reinforcementlearning)
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
DeepLearningFrameworks
A
B
C
Pytorch
- Dynamicgraph사용
- 학계에서많이사용
Tensorflow
- Staticgraph사용
- 산업체에서많이사용
etc
- Caffe,DEEPLEARNING4J,...
- 최근에는Pytorch와Tensorflow양강체제
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
DeepLearning기본단어
epoch, batch, loss, optimizer, back propagation, ...
Name Meaning
epoch 큰trainingloop의단위
(mini)batch Training할때model이한번에보는양
Loss
최소화하고자하는대상
Network는lossfunction의globalminima를찾기위해training함
Optimizer Loss가globalminima를효과적으로찾을수있도록도움을주는것
Learningrate Update된gradient의반영비율
Back
propagation
Network의weigh를update시키는방법
activation 일반적으로non-linearfunction을가리킴
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
DeepLearning기본단어
epoch, batch, loss, optimizer, back propagation, ...
Name Meaning
epoch 큰trainingloop의단위
(mini)batch Training할때model이한번에보는양
Loss
최소화하고자하는대상
Network는lossfunction의globalminima를찾기위해training함
Optimizer Loss가globalminima를효과적으로찾을수있도록도움을주는것
Learningrate Update된gradient의반영비율
Back
propagation
Network의weigh를update시키는방법
activation 일반적으로non-linearfunction을가리킴
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
출처: https://www.kakaobrain.com/blog/113
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
출처: https://gaussian37.github.io/dl-concept-batchnorm/
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
DeepLearning기본단어
epoch, batch, loss, optimizer, back propagation, ...
Name Meaning
epoch 큰trainingloop의단위
(mini)batch Training할때model이한번에보는양
Loss
최소화하고자하는대상
Network는lossfunction의globalminima를찾기위해training함
Optimizer Loss가globalminima를효과적으로찾을수있도록도움을주는것
Learningrate Update된gradient의반영비율
Back
propagation
Network의weigh를update시키는방법
activation 일반적으로non-linearfunction을가리킴
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
출처: https://financial-engineering.medium.com/tensorflow-2-0-loss-function-%EC%86%90%EC%8B%A4%ED%95%A8%EC%88%98-a01b0a1492a7
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
출처: https://laptrinhx.com/a-way-to-improve-gradient-descent-stochastic-gradient-descent-with-restarts-3873821469/
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
출처: https://laptrinhx.com/a-way-to-improve-gradient-descent-stochastic-gradient-descent-with-restarts-3873821469/
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
DeepLearning기본단어
epoch, batch, loss, optimizer, back propagation, ...
Name Meaning
epoch 큰trainingloop의단위
(mini)batch Training할때model이한번에보는양
Loss
최소화하고자하는대상
Network는lossfunction의globalminima를찾기위해training함
Optimizer Loss가globalminima를효과적으로찾을수있도록도움을주는것
Learningrate Update된gradient의반영비율
Back
propagation
Network의weigh를update시키는방법
activation 일반적으로non-linearfunction을가리킴
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
출처: https://ganghee-lee.tistory.com/24
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
DeepLearning기본단어
epoch, batch, loss, optimizer, back propagation, ...
Name Meaning
epoch 큰trainingloop의단위
(mini)batch Training할때model이한번에보는양
Loss
최소화하고자하는대상
Network는lossfunction의globalminima를찾기위해training함
Optimizer Loss가globalminima를효과적으로찾을수있도록도움을주는것
Learningrate Update된gradient의반영비율
Back
propagation
Network의weigh를update시키는방법
activation 일반적으로non-linearfunction을가리킴
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
출처: https://www.jeremyjordan.me/nn-learning-rate/
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
출처: https://medium.com/analytics-vidhya/this-blog-post-aims-at-explaining-the-behavior-of-different-algorithms-for-optimizing-gradient-46159a97a8c1
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
DeepLearning기본단어
epoch, batch, loss, optimizer, back propagation, ...
Name Meaning
epoch 큰trainingloop의단위
(mini)batch Training할때model이한번에보는양
Loss
최소화하고자하는대상
Network는lossfunction의globalminima를찾기위해training함
Optimizer Loss가globalminima를효과적으로찾을수있도록도움을주는것
Learningrate Update된gradient의반영비율
Back
propagation
Network의weight를update시키는방법
activation 일반적으로non-linearfunction을가리킴
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
출처: https://gjghks.tistory.com/75
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
DeepLearning기본단어
epoch, batch, loss, optimizer, back propagation, ...
Name Meaning
epoch 큰trainingloop의단위
(mini)batch Training할때model이한번에보는양
Loss
최소화하고자하는대상
Network는lossfunction의globalminima를찾기위해training함
Optimizer Loss가globalminima를효과적으로찾을수있도록도움을주는것
Learningrate Update된gradient의반영비율
Back
propagation
Network의weigh를update시키는방법
activation 일반적으로non-linearfunction을가리킴
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
출처: https://medium.com/@shrutijadon10104776/survey-on-activation-functions-for-deep-learning-9689331ba092
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
Part 2,
CNN
(ConvolutionalNeuralNetwork)
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
FromFCtoConv
왜 우리는 convolutional layer를 사용해야 하는가
공간정보 X
Fully Connected
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
Q.
공간정보는어떻게보존하죠?
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
#Convolution
이에대한해결책은...
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
FromFCtoConv
왜 우리는 convolutional layer를 사용해야 하는가
공간정보 X
Fully Connected
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
FromFCtoConv
왜 우리는 convolutional layer를 사용해야 하는가
공간정보 X
공간정보 O
Fully Connected Convolution
출처: https://mlnotebook.github.io/post/CNN1/
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
FromFCtoConv
왜 우리는 convolutional layer를 사용해야 하는가
공간정보 X
공간정보 O
Fully Connected Convolution
출처: https://mlnotebook.github.io/post/CNN1/
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
CNNprocess
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
CNNprocess
Layer
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
CNNprocess
Layer
출처: https://towardsdatascience.com/understanding-1d-and-3d-convolution-neural-network-keras-9d8f76e29610
[batch, height, width, channel]
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
Kernel(filter),stride,weight,bias
kernel: [k_x, k_y, ch_in, ch_out]
출처: http://taewan.kim/post/cnn/
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
Padding
feature map size가 줄어들지 않도록...
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
Pooling
feature map size를 줄이거나 특정 데이터를 강조하고 싶다면?
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
ReceptiveField
하나의 output point를 내기 위해 필요한 input range는?
출처: https://www.researchgate.net/figure/The-receptive-field-of-each-convolution-
layer-with-a-3-3-kernel-The-green-area-marks_fig4_316950618
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
Sub-pixelconvolution
Super resolution이 일어나는 convolution
출처: https://medium.com/@zhuocen93/an-overview-of-espcn-an-efficient-sub-pixel-convolutional-neural-
network-b76d0a6c875e
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
PopularCNNmodels
VGG, ResNet, DenseNet, ...
VGG
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
PopularCNNmodels
VGG, ResNet, DenseNet, ...
ResNet
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
PopularCNNmodels
VGG, ResNet, DenseNet, ...
DenseNet
출처: https://ai-pool.com/m/densenet-1568742493
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
Part 3,
GAN
(GenerativeAdversarialNetwork)
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
GAN기본개념
경찰과 위조지폐범의 숨막히는 결전
출처: https://m.blog.naver.com/euleekwon/221557899873
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
GANexamples
Fake Obama: 2017 워싱턴대학교
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
GANexamples
AI 음악 프로젝트: 다시 한 번 – AI 거북이
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
GANexamples
ESRGAN
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
GANlosses
adversarial loss, content loss
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
GANmodel의장단점
무에서 유를 창조해내는 힘 vs 그게 그렇게 쉽게 될리가
- MSEloss로는얻어낼수없는결과를만들어냄
장점:압도적인성능
- generator와discriminator와의균형을맞추기가어려움
- modecollapse등의문제가발생할우려가있음
단점:training이어려움
ⓒSaebyeol Yu. Saebyeol’s PowerPoint
감사합니다

More Related Content

PDF
오토인코더의 모든 것
PDF
Recent advances of AI for medical imaging : Engineering perspectives
PDF
[2A4]DeepLearningAtNAVER
PPTX
White box in Computer Vision
PDF
Applying deep learning to medical data
PPTX
Deep Learning for AI (2)
PPTX
AutoEncoder&GAN Introduction
PPTX
Deep Learning for AI (3)
오토인코더의 모든 것
Recent advances of AI for medical imaging : Engineering perspectives
[2A4]DeepLearningAtNAVER
White box in Computer Vision
Applying deep learning to medical data
Deep Learning for AI (2)
AutoEncoder&GAN Introduction
Deep Learning for AI (3)

Similar to Introduction to deep learning (20)

PDF
Introduction to GAN
PDF
Deep learning image recognition for autonomous driving(classification, objec...
PDF
2024_개보위_개인정보 미래포럼_의료 인공지능 모델과 프라이버시 이슈.pdf
PDF
Matching Network
PDF
Lecture 5: Neural Networks II
PPTX
Convolutional neural networks 이론과 응용
PDF
Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning
PDF
ujava.org Deep Learning with Convolutional Neural Network
PDF
Koss 1605 machine_learning_mariocho_t10
PDF
Nlp and transformer (v3s)
PDF
Deep learning for Junior Developer
PPTX
Graph Neural Network #2-1 (PinSage)
PDF
ESM Machine learning 5주차 Review by Mario Cho
PPTX
PPT - Enhancing the Locality and Breaking the Memory Bottleneck of Transforme...
PPTX
20230213_ComputerVision_연구.pptx
PPT
Deep learning for Junior Developer
PDF
H transformer-1d paper review!!
PDF
[PR12] understanding deep learning requires rethinking generalization
PDF
알고리즘 중심의 머신러닝 가이드 Ch04
PDF
Photo-realistic Single Image Super-resolution using a Generative Adversarial ...
Introduction to GAN
Deep learning image recognition for autonomous driving(classification, objec...
2024_개보위_개인정보 미래포럼_의료 인공지능 모델과 프라이버시 이슈.pdf
Matching Network
Lecture 5: Neural Networks II
Convolutional neural networks 이론과 응용
Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning
ujava.org Deep Learning with Convolutional Neural Network
Koss 1605 machine_learning_mariocho_t10
Nlp and transformer (v3s)
Deep learning for Junior Developer
Graph Neural Network #2-1 (PinSage)
ESM Machine learning 5주차 Review by Mario Cho
PPT - Enhancing the Locality and Breaking the Memory Bottleneck of Transforme...
20230213_ComputerVision_연구.pptx
Deep learning for Junior Developer
H transformer-1d paper review!!
[PR12] understanding deep learning requires rethinking generalization
알고리즘 중심의 머신러닝 가이드 Ch04
Photo-realistic Single Image Super-resolution using a Generative Adversarial ...
Ad

Recently uploaded (20)

PPTX
CARTOGRAPHY AND GEOINFORMATION VISUALIZATION chapter1 NPTE (2).pptx
PDF
Mitigating Risks through Effective Management for Enhancing Organizational Pe...
PPTX
MET 305 2019 SCHEME MODULE 2 COMPLETE.pptx
PPTX
Artificial Intelligence
PDF
keyrequirementskkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkk
PDF
composite construction of structures.pdf
PPTX
FINAL REVIEW FOR COPD DIANOSIS FOR PULMONARY DISEASE.pptx
PPTX
Current and future trends in Computer Vision.pptx
PDF
Mohammad Mahdi Farshadian CV - Prospective PhD Student 2026
PDF
R24 SURVEYING LAB MANUAL for civil enggi
PPTX
Engineering Ethics, Safety and Environment [Autosaved] (1).pptx
PPTX
Safety Seminar civil to be ensured for safe working.
DOCX
573137875-Attendance-Management-System-original
PPTX
M Tech Sem 1 Civil Engineering Environmental Sciences.pptx
PDF
Automation-in-Manufacturing-Chapter-Introduction.pdf
PPT
Project quality management in manufacturing
PPTX
Construction Project Organization Group 2.pptx
PPTX
additive manufacturing of ss316l using mig welding
PPTX
OOP with Java - Java Introduction (Basics)
PDF
III.4.1.2_The_Space_Environment.p pdffdf
CARTOGRAPHY AND GEOINFORMATION VISUALIZATION chapter1 NPTE (2).pptx
Mitigating Risks through Effective Management for Enhancing Organizational Pe...
MET 305 2019 SCHEME MODULE 2 COMPLETE.pptx
Artificial Intelligence
keyrequirementskkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkk
composite construction of structures.pdf
FINAL REVIEW FOR COPD DIANOSIS FOR PULMONARY DISEASE.pptx
Current and future trends in Computer Vision.pptx
Mohammad Mahdi Farshadian CV - Prospective PhD Student 2026
R24 SURVEYING LAB MANUAL for civil enggi
Engineering Ethics, Safety and Environment [Autosaved] (1).pptx
Safety Seminar civil to be ensured for safe working.
573137875-Attendance-Management-System-original
M Tech Sem 1 Civil Engineering Environmental Sciences.pptx
Automation-in-Manufacturing-Chapter-Introduction.pdf
Project quality management in manufacturing
Construction Project Organization Group 2.pptx
additive manufacturing of ss316l using mig welding
OOP with Java - Java Introduction (Basics)
III.4.1.2_The_Space_Environment.p pdffdf
Ad

Introduction to deep learning

  • 1. ⓒSaebyeol Yu. Saebyeol’s PowerPoint Hello, DL! Introductiontodeeplearning ⓒSaebyeol Yu. Saebyeol’s PowerPoint ⓒSaebyeol Yu. Saebyeol’s PowerPoint Robin
  • 2. ⓒSaebyeol Yu. Saebyeol’s PowerPoint a table of contents 목차 1 Deeplearning이란? 2 CNN 3 GAN
  • 3. ⓒSaebyeol Yu. Saebyeol’s PowerPoint Part 1, DeepLearning이란?
  • 4. ⓒSaebyeol Yu. Saebyeol’s PowerPoint DeepLearning이란? MachineLearning
  • 5. ⓒSaebyeol Yu. Saebyeol’s PowerPoint DeepLearning이란? MachineLearning Neuron
  • 6. ⓒSaebyeol Yu. Saebyeol’s PowerPoint DeepLearning이란? MachineLearning Neuron DeepLearning
  • 7. ⓒSaebyeol Yu. Saebyeol’s PowerPoint DeepLearning이란? 출처: https://blog.naver.com/PostView.nhn?blogId=dsjang650628&logNo=221864626337
  • 8. ⓒSaebyeol Yu. Saebyeol’s PowerPoint DeepLearning의종류 Network schemes 에 따른 분류 CNN (ConvolutionalNeuralNetwork) GAN (GenerativeAdversarialNetwork) RNN (RecurrentNeuralNetwork) LSTM (LongShort-TermMemorymodels)
  • 9. ⓒSaebyeol Yu. Saebyeol’s PowerPoint DeepLearning의종류 Applications 에 따른 분류 CV (ComputerVision) NLP (NaturalLanguageProcessing) RF (Reinforcementlearning)
  • 10. ⓒSaebyeol Yu. Saebyeol’s PowerPoint DeepLearningFrameworks A B C Pytorch - Dynamicgraph사용 - 학계에서많이사용 Tensorflow - Staticgraph사용 - 산업체에서많이사용 etc - Caffe,DEEPLEARNING4J,... - 최근에는Pytorch와Tensorflow양강체제
  • 11. ⓒSaebyeol Yu. Saebyeol’s PowerPoint DeepLearning기본단어 epoch, batch, loss, optimizer, back propagation, ... Name Meaning epoch 큰trainingloop의단위 (mini)batch Training할때model이한번에보는양 Loss 최소화하고자하는대상 Network는lossfunction의globalminima를찾기위해training함 Optimizer Loss가globalminima를효과적으로찾을수있도록도움을주는것 Learningrate Update된gradient의반영비율 Back propagation Network의weigh를update시키는방법 activation 일반적으로non-linearfunction을가리킴
  • 12. ⓒSaebyeol Yu. Saebyeol’s PowerPoint DeepLearning기본단어 epoch, batch, loss, optimizer, back propagation, ... Name Meaning epoch 큰trainingloop의단위 (mini)batch Training할때model이한번에보는양 Loss 최소화하고자하는대상 Network는lossfunction의globalminima를찾기위해training함 Optimizer Loss가globalminima를효과적으로찾을수있도록도움을주는것 Learningrate Update된gradient의반영비율 Back propagation Network의weigh를update시키는방법 activation 일반적으로non-linearfunction을가리킴
  • 13. ⓒSaebyeol Yu. Saebyeol’s PowerPoint 출처: https://www.kakaobrain.com/blog/113
  • 14. ⓒSaebyeol Yu. Saebyeol’s PowerPoint 출처: https://gaussian37.github.io/dl-concept-batchnorm/
  • 15. ⓒSaebyeol Yu. Saebyeol’s PowerPoint DeepLearning기본단어 epoch, batch, loss, optimizer, back propagation, ... Name Meaning epoch 큰trainingloop의단위 (mini)batch Training할때model이한번에보는양 Loss 최소화하고자하는대상 Network는lossfunction의globalminima를찾기위해training함 Optimizer Loss가globalminima를효과적으로찾을수있도록도움을주는것 Learningrate Update된gradient의반영비율 Back propagation Network의weigh를update시키는방법 activation 일반적으로non-linearfunction을가리킴
  • 16. ⓒSaebyeol Yu. Saebyeol’s PowerPoint 출처: https://financial-engineering.medium.com/tensorflow-2-0-loss-function-%EC%86%90%EC%8B%A4%ED%95%A8%EC%88%98-a01b0a1492a7
  • 17. ⓒSaebyeol Yu. Saebyeol’s PowerPoint 출처: https://laptrinhx.com/a-way-to-improve-gradient-descent-stochastic-gradient-descent-with-restarts-3873821469/
  • 18. ⓒSaebyeol Yu. Saebyeol’s PowerPoint 출처: https://laptrinhx.com/a-way-to-improve-gradient-descent-stochastic-gradient-descent-with-restarts-3873821469/
  • 19. ⓒSaebyeol Yu. Saebyeol’s PowerPoint DeepLearning기본단어 epoch, batch, loss, optimizer, back propagation, ... Name Meaning epoch 큰trainingloop의단위 (mini)batch Training할때model이한번에보는양 Loss 최소화하고자하는대상 Network는lossfunction의globalminima를찾기위해training함 Optimizer Loss가globalminima를효과적으로찾을수있도록도움을주는것 Learningrate Update된gradient의반영비율 Back propagation Network의weigh를update시키는방법 activation 일반적으로non-linearfunction을가리킴
  • 20. ⓒSaebyeol Yu. Saebyeol’s PowerPoint 출처: https://ganghee-lee.tistory.com/24
  • 21. ⓒSaebyeol Yu. Saebyeol’s PowerPoint DeepLearning기본단어 epoch, batch, loss, optimizer, back propagation, ... Name Meaning epoch 큰trainingloop의단위 (mini)batch Training할때model이한번에보는양 Loss 최소화하고자하는대상 Network는lossfunction의globalminima를찾기위해training함 Optimizer Loss가globalminima를효과적으로찾을수있도록도움을주는것 Learningrate Update된gradient의반영비율 Back propagation Network의weigh를update시키는방법 activation 일반적으로non-linearfunction을가리킴
  • 22. ⓒSaebyeol Yu. Saebyeol’s PowerPoint 출처: https://www.jeremyjordan.me/nn-learning-rate/
  • 23. ⓒSaebyeol Yu. Saebyeol’s PowerPoint 출처: https://medium.com/analytics-vidhya/this-blog-post-aims-at-explaining-the-behavior-of-different-algorithms-for-optimizing-gradient-46159a97a8c1
  • 24. ⓒSaebyeol Yu. Saebyeol’s PowerPoint DeepLearning기본단어 epoch, batch, loss, optimizer, back propagation, ... Name Meaning epoch 큰trainingloop의단위 (mini)batch Training할때model이한번에보는양 Loss 최소화하고자하는대상 Network는lossfunction의globalminima를찾기위해training함 Optimizer Loss가globalminima를효과적으로찾을수있도록도움을주는것 Learningrate Update된gradient의반영비율 Back propagation Network의weight를update시키는방법 activation 일반적으로non-linearfunction을가리킴
  • 25. ⓒSaebyeol Yu. Saebyeol’s PowerPoint 출처: https://gjghks.tistory.com/75
  • 26. ⓒSaebyeol Yu. Saebyeol’s PowerPoint DeepLearning기본단어 epoch, batch, loss, optimizer, back propagation, ... Name Meaning epoch 큰trainingloop의단위 (mini)batch Training할때model이한번에보는양 Loss 최소화하고자하는대상 Network는lossfunction의globalminima를찾기위해training함 Optimizer Loss가globalminima를효과적으로찾을수있도록도움을주는것 Learningrate Update된gradient의반영비율 Back propagation Network의weigh를update시키는방법 activation 일반적으로non-linearfunction을가리킴
  • 27. ⓒSaebyeol Yu. Saebyeol’s PowerPoint 출처: https://medium.com/@shrutijadon10104776/survey-on-activation-functions-for-deep-learning-9689331ba092
  • 28. ⓒSaebyeol Yu. Saebyeol’s PowerPoint Part 2, CNN (ConvolutionalNeuralNetwork)
  • 29. ⓒSaebyeol Yu. Saebyeol’s PowerPoint FromFCtoConv 왜 우리는 convolutional layer를 사용해야 하는가 공간정보 X Fully Connected
  • 30. ⓒSaebyeol Yu. Saebyeol’s PowerPoint Q. 공간정보는어떻게보존하죠?
  • 31. ⓒSaebyeol Yu. Saebyeol’s PowerPoint #Convolution 이에대한해결책은...
  • 32. ⓒSaebyeol Yu. Saebyeol’s PowerPoint FromFCtoConv 왜 우리는 convolutional layer를 사용해야 하는가 공간정보 X Fully Connected
  • 33. ⓒSaebyeol Yu. Saebyeol’s PowerPoint FromFCtoConv 왜 우리는 convolutional layer를 사용해야 하는가 공간정보 X 공간정보 O Fully Connected Convolution 출처: https://mlnotebook.github.io/post/CNN1/
  • 34. ⓒSaebyeol Yu. Saebyeol’s PowerPoint FromFCtoConv 왜 우리는 convolutional layer를 사용해야 하는가 공간정보 X 공간정보 O Fully Connected Convolution 출처: https://mlnotebook.github.io/post/CNN1/
  • 35. ⓒSaebyeol Yu. Saebyeol’s PowerPoint CNNprocess
  • 36. ⓒSaebyeol Yu. Saebyeol’s PowerPoint CNNprocess Layer
  • 37. ⓒSaebyeol Yu. Saebyeol’s PowerPoint CNNprocess Layer 출처: https://towardsdatascience.com/understanding-1d-and-3d-convolution-neural-network-keras-9d8f76e29610 [batch, height, width, channel]
  • 38. ⓒSaebyeol Yu. Saebyeol’s PowerPoint Kernel(filter),stride,weight,bias kernel: [k_x, k_y, ch_in, ch_out] 출처: http://taewan.kim/post/cnn/
  • 39. ⓒSaebyeol Yu. Saebyeol’s PowerPoint Padding feature map size가 줄어들지 않도록...
  • 40. ⓒSaebyeol Yu. Saebyeol’s PowerPoint Pooling feature map size를 줄이거나 특정 데이터를 강조하고 싶다면?
  • 41. ⓒSaebyeol Yu. Saebyeol’s PowerPoint ReceptiveField 하나의 output point를 내기 위해 필요한 input range는? 출처: https://www.researchgate.net/figure/The-receptive-field-of-each-convolution- layer-with-a-3-3-kernel-The-green-area-marks_fig4_316950618
  • 42. ⓒSaebyeol Yu. Saebyeol’s PowerPoint Sub-pixelconvolution Super resolution이 일어나는 convolution 출처: https://medium.com/@zhuocen93/an-overview-of-espcn-an-efficient-sub-pixel-convolutional-neural- network-b76d0a6c875e
  • 43. ⓒSaebyeol Yu. Saebyeol’s PowerPoint PopularCNNmodels VGG, ResNet, DenseNet, ... VGG
  • 44. ⓒSaebyeol Yu. Saebyeol’s PowerPoint PopularCNNmodels VGG, ResNet, DenseNet, ... ResNet
  • 45. ⓒSaebyeol Yu. Saebyeol’s PowerPoint PopularCNNmodels VGG, ResNet, DenseNet, ... DenseNet 출처: https://ai-pool.com/m/densenet-1568742493
  • 46. ⓒSaebyeol Yu. Saebyeol’s PowerPoint Part 3, GAN (GenerativeAdversarialNetwork)
  • 47. ⓒSaebyeol Yu. Saebyeol’s PowerPoint GAN기본개념 경찰과 위조지폐범의 숨막히는 결전 출처: https://m.blog.naver.com/euleekwon/221557899873
  • 48. ⓒSaebyeol Yu. Saebyeol’s PowerPoint GANexamples Fake Obama: 2017 워싱턴대학교
  • 49. ⓒSaebyeol Yu. Saebyeol’s PowerPoint GANexamples AI 음악 프로젝트: 다시 한 번 – AI 거북이
  • 50. ⓒSaebyeol Yu. Saebyeol’s PowerPoint GANexamples ESRGAN
  • 51. ⓒSaebyeol Yu. Saebyeol’s PowerPoint GANlosses adversarial loss, content loss
  • 52. ⓒSaebyeol Yu. Saebyeol’s PowerPoint GANmodel의장단점 무에서 유를 창조해내는 힘 vs 그게 그렇게 쉽게 될리가 - MSEloss로는얻어낼수없는결과를만들어냄 장점:압도적인성능 - generator와discriminator와의균형을맞추기가어려움 - modecollapse등의문제가발생할우려가있음 단점:training이어려움
  • 53. ⓒSaebyeol Yu. Saebyeol’s PowerPoint 감사합니다