Changwon National University, profile picture
Uploaded byChangwon National University
PDF, PPTX340 views

AI 로봇 아티스트의 비밀(창원대학교 정보통신공학과 특강)

The document discusses deep generative models in artificial intelligence, focusing on techniques such as autoencoders, variational autoencoders (VAEs), and generative adversarial networks (GANs). It explains the concepts of supervised and unsupervised learning, generative modeling, and the purpose of latent variables in creating representations of data. Key insights include understanding how these models learn from data to generate new samples and improve their performance in various applications.

Embed presentation

Download as PDF, PPTX
1 / 102
2 / 102
3 / 102
4 / 102
5 / 102
6 / 102
7 / 102
8 / 102
9 / 102
10 / 102
11 / 102
12 / 102
13 / 102
14 / 102
15 / 102
16 / 102
17 / 102
18 / 102
19 / 102
20 / 102
21 / 102
22 / 102
23 / 102
24 / 102
25 / 102
26 / 102
27 / 102
28 / 102
29 / 102
30 / 102
31 / 102
32 / 102
33 / 102
34 / 102
35 / 102
36 / 102
37 / 102
38 / 102
39 / 102
40 / 102
41 / 102
42 / 102
43 / 102
44 / 102
45 / 102
46 / 102
47 / 102
48 / 102
49 / 102
50 / 102
51 / 102
52 / 102
53 / 102
54 / 102
55 / 102
56 / 102
57 / 102
58 / 102
59 / 102
60 / 102
61 / 102
62 / 102
63 / 102
64 / 102
65 / 102
66 / 102
67 / 102
68 / 102
69 / 102
70 / 102
71 / 102
72 / 102
73 / 102
74 / 102
75 / 102
76 / 102
77 / 102
78 / 102
79 / 102
80 / 102
81 / 102
82 / 102
83 / 102
84 / 102
85 / 102
86 / 102
87 / 102
88 / 102
89 / 102
90 / 102
91 / 102
92 / 102
93 / 102
94 / 102
95 / 102
96 / 102
97 / 102
98 / 102
99 / 102
100 / 102
101 / 102
102 / 102
AI 로봇 아티스트의 비밀 Deep Generative Models 2019-05-23 Sung-Soo Kim sungsoo@etri.re.kr Smart Data Research Group 특강
Outline ! Introduction – ArtificialIntelligence,MachineLearning,DeepLearning ! Autoencoders ! Variational Autoencoders (VAEs) ! Generative Adversarial Networks (GANs) ! Summary
Slide Credits
4 6.S191 Introduction to Deep Learning Which face is fake?
AI Artist: Ai-Da
6 ArtificialIntelligence MachineLearning
인공지능이란 무엇일까요?
인공지능을 단순화해서 생각해봅시다.
Dog Cat 입력이 주어지면 출력을 내보낸다. 인공지능
입력이 주어지면, 출력을 내보낸다. 우리 이런 걸 어디서 봤죠?
함수 (Functions) System, Filter 라고도 불립니다.
인공지능 별 거 아니죠? 무서울 것 없습니다. 근데 왜 갑자기 무서워 졌을까요?
모두의연구소 수학적으로 표현할 수 없었던 복잡한 인간의 두뇌를 데이터를 기반으로 흉내내다.
러닝(Learning) = 학습 Adaptation/Update
러닝(Learning) = 학습 Adaptation/Update 인공지능이 점점 똑똑해진다.
X[n] F[n] F[n-1] Y[n] n : data index Data가 늘어날수록 점점 인공지능 알고리즘이 학습(Learning)한다.
머신러닝의 세가지 타입 Supervised Unsupervised Reinforcement Learning Classification Regression Clustering
3 X ??? 6 Dog Cat Supervised Learning - 정답이 주어진다. - (비교적)문제풀이가 쉽다
3 X ??? ??? Unsupervised Learning 미지수 2개, 방정식 1개 풀 수 있나요?
Reinforcement Learning - (정답이 아닌) reward가 주어진다. Reward
x ??? +10 / -10 Reinforcement Learning
딥러닝을 이해하기 위해서는 인공신경망을 알아야 한다.
뉴런과 인공뉴런 x 1 x 1 b a 다른 머신러닝 기법들과의 차이점 1: Nonlinear(복잡한) activation function b a
다층 레이어 (Multiple Layer) 다른 머신러닝 기법들과의 차이점 2 : Nonlinear function의 Nonlinear function의 Nonlinear function … 엄청 복잡한 함수 (인공지능)을 만들 수 있다. Hidden layer가 2개이상인 NN를 Deep Learning이라고 부른다.
머신러닝 (Machine Learning) 인공신경망 (Neural Network) 딥러닝 (Deep Learning) 딥러닝? 머신러닝?
26 6.S191 Introduction to Deep Learning Supervised vs unsupervised learning Supervised Learning Data: (", $) " is data, $ is label Goal: Learn function to map " → $ Examples: Classification, regression, object detection, semantic segmentation, etc. Unsupervised Learning Data: " " is data, no labels! Goal: Learn some hidden or underlying structure of the data Examples: Clustering, feature or dimensionality reduction, etc.
27 6.S191 Introduction to Deep Learning Supervised vs unsupervised learning Supervised Learning Data: (", $) " is data, $ is label Goal: Learn function to map " → $ Examples: Classification, regression, object detection, semantic segmentation, etc. Unsupervised Learning Data: " " is data, no labels! Goal: Learn some hidden or underlying structure of the data Examples: Clustering, feature or dimensionality reduction, etc.
28 6.S191 Introduction to Deep Learning Generative modeling Goal: Take as input training samples from some distribution and learn a model that represents that distribution Density Estimation Sample Generation Input samples Generated samples Training data ~ "#$%$ & Generated ~ "'(#)* & How can we learn "'(#)* & similar to "#$%$ & ? samples KeyTakeaway#1 KeyTakeaway#2 MainTakeaways
29 6.S191 Introduction to Deep Learning Why generative models? Debiasing vs Capable of uncovering underlying latent variables in a dataset Homogeneous skin color, pose Diverse skin color, pose, illumination How can we use latent distributions to create fair and representative datasets?
30 6.S191 Introduction to Deep Learning Why generative models? Outlier detection 95% of Driving Data: (1) sunny, (2) highway, (3) straight road Detect outliers to avoid unpredictable behavior when training Edge Cases Harsh Weather Pedestrians • Problem: How can we detect when we encounter something new or rare? • Strategy: Leverage generative models, detect outliers in the distribution • Use outliers during training to improve even more!
31 Latent variable models Autoencoders andVariational Autoencoders (VAEs) Generative Adversarial Networks (GANs)
32 6.S191 Introduction to Deep Learning What is a latent variable? Myth of the Cave 515c~d
33 Matrix (1999)
34 6.S191 Introduction to Deep Learning What is a latent variable? Can we learn the true explanatory factors, e.g. latent variables, from only observed data?
35 Autoencoders
36 6.S191 Introduction to Deep Learning Autoencoders: background Unsupervised approach for learning a lower-dimensional feature representation from unlabeled training data ! " “Encoder” learns mapping from the data, !, to a low-dimensional latent space, " Why do we care about a low-dimensional "?
37 6.S191 Introduction to Deep Learning Autoencoders: background How can we learn this latent space? Train the model to use these features to reconstruct the original data ! " “Decoder” learns mapping back from latent, ", to a reconstructed observation, #! #!
38 6.S191 Introduction to Deep Learning Autoencoders: background How can we learn this latent space? Train the model to use these features to reconstruct the original data ! " #! ℒ !, #! = ! − #! ( Loss function doesn’t use any labels!!
39 6.S191 Introduction to Deep Learning Autoencoders: background How can we learn this latent space? Train the model to use these features to reconstruct the original data ! " #! ℒ !, #! = ! − #! ( Loss function doesn’t use any labels!!
40 6.S191 Introduction to Deep Learning Dimensionality of latent space à reconstruction quality 2D latent space 5D latent space GroundTruth Autoencoding is a form of compression! Smaller latent space will force a larger training bottleneck
41 6.S191 Introduction to Deep Learning Autoencoders for representation learning Bottleneck hidden layer forces network to learn a compressed latent representation Reconstruction loss forces the latent representation to capture (or encode) as much “information” about the data as possible Autoencoding = Automatically encoding data
42 Variational Autoencoders (VAEs)
43 2019 ModuLabs Taxonomy of Generative Models Figure credit: Ian Goodfellow, NIPS 2016 Tutorial: Generative Adversarial Networks
44 6.S191 Introduction to Deep Learning VAEs: key difference with traditional autoencoder ! " #!
45 6.S191 Introduction to Deep Learning VAEs: key difference with traditional autoencoder ! " #! $ %
46 6.S191 Introduction to Deep Learning VAEs: key difference with traditional autoencoder ! " #! $ % mean vector standard deviation vector Variational autoencoders are a probabilistic twist on autoencoders! Sample from the mean and standard dev. to compute latent sample
47 6.S191 Introduction to Deep Learning VAE optimization ! " #! $ % Encoder computes: &'(z|!) Decoder computes: ,-(x|")
48 6.S191 Introduction to Deep Learning VAE optimization ! " #! $ % Encoder computes: &'(z|!) Decoder computes:,-(x|") ℒ ϕ, 2 = (reconstruction loss) + (regularization term)
49 6.S191 Introduction to Deep Learning VAE optimization ! " #! $ % Encoder computes: &'(z|!) Decoder computes:,-(x|") ℒ ϕ, 2, ! = (reconstruction loss) + (regularization term)
50 6.S191 Introduction to Deep Learning VAE optimization ! " #! $ % Encoder computes: &'(z|!) Decoder computes:,-(x|") ℒ ϕ, 2, ! = (reconstruction loss) + (regularization term) e.g. ! − #! 5
51 6.S191 Introduction to Deep Learning VAE optimization ! " #! $ % Encoder computes: &'(z|!) Decoder computes:,-(x|") ℒ ϕ, 2, ! = (reconstruction loss) + (regularization term) 4 &' z ! ∥ & " Inferred latent distribution Fixed prior on latent distribution
52 6.S191 Introduction to Deep Learning Priors on the latent distribution ! "# z % ∥ " ' Inferred latent distribution Fixed prior on latent distribution Common choice of prior: " ' = ) * = 0, -. = 1 • Encourages encodings to distribute encodings evenly around the center of the latent space • Penalize the network when it tries to “cheat” by clustering points in specific regions (ie. memorizing the data)
53 6.S191 Introduction to Deep Learning Priors on the latent distribution ! "# z % ∥ " ' Common choice of prior: " ' = ) * = 0, -. = 1 • Encourages encodings to distribute encodings evenly around the center of the latent space • Penalize the network when it tries to “cheat” by clustering points in specific regions (ie. memorizing the data) = − 1 2 2 345 678 -3 + *3 . − 1 − log -3 KL-divergence between the two distributions
54 6.S191 Introduction to Deep Learning VAEs computation graph ! " #! $ % Encoder computes: &'(z|!) Decoder computes:,-(x|") ℒ ϕ, 2, ! = (reconstruction loss) + (regularization term)
55 6.S191 Introduction to Deep Learning VAEs computation graph ! " #! $ % Encoder computes: &'(z|!) Decoder computes:,-(x|") ℒ ϕ, 2, ! = (reconstruction loss) + (regularization term) Problem: We cannot backpropagate gradients through sampling layers!
56 6.S191 Introduction to Deep Learning Reparametrizing the sampling layer ! " # Key Idea: ! ~%(", #() Consider the sampled latent vector as a sum of • a fixed " vector, • and fixed # vector, scaled by random constants drawn from the prior distribution ⇒ ! = " + #⨀. where .~%(0,1)
57 6.S191 Introduction to Deep Learning Reparametrizing the sampling layer ! "" ∼ $%(z|)) + ) Deterministic node Stochastic node Original form Backprop
58 6.S191 Introduction to Deep Learning Reparametrizing the sampling layer ! "" ∼ $%(z|)) ! " " = ,(-, ), /) - ) /- ) Deterministic node Stochastic node Original form Reparametrized form 0! 0" 0! 0- Backprop ~2(0,1)
59 6.S191 Introduction to Deep Learning VAEs: Latent perturbation Slowly increase or decrease a single latent variable Keep all other variables fixed Head pose Different dimensions of ! encodes different interpretable latent features
60 6.S191 Introduction to Deep Learning VAEs: Latent perturbation Head pose Smile Ideally, we want latent variables that are uncorrelated with each other Enforce diagonal prior on the latent variables to encourage independence Disentanglement
61 6.S191 Introduction to Deep Learning VAEs: Latent perturbation Google BeatBlender GoogleBeatBlender
62 GoogleBeatBlender
63 6.S191 Introduction to Deep Learning VAE summary 1. Compress representation of world to something we can use to learn 2. Reconstruction allows for unsupervised learning (no labels!) 3. Reparameterization trick to train end-to-end 4. Interpret hidden latent variables using perturbation 5. Generating new examples ! " #!
64 Generative Adversarial Networks (GANs)
65 2019 ModuLabs “The most important idea in Machine Learning in the last 20 years”
66 2019 ModuLabs Taxonomy of Generative Models Figure credit: Ian Goodfellow, NIPS 2016 Tutorial: Generative Adversarial Networks
67 6.S191 Introduction to Deep Learning What if we just want to sample? Idea: don’t explicitly model density, and instead just sample to generate new instances. Problem: want to sample from complex distribution – can’t do this directly! Solution: sample from something simple (noise), learn a transformation to the training distribution. !noise " #$%&' “fake” sample from the training distribution Generator Network "
68 6.S191 Introduction to Deep Learning Generative Adversarial Networks (GANs) Generative Adversarial Networks (GANs) are a way to make a generative model by having two neural networks compete with each other. The discriminator tries to identify real data from fakes created by the generator. The generator turns noise into an imitation of the data to try to trick the discriminator. !noise " # $%&'($)'*& +
69 Intuition
70 GameTheory and GANs
71 6.S191 Introduction to Deep Learning Intuition behind GANs Generator Generator starts from noise to try to create an imitation of the data. Fake data OneMoreTime!
72 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator Generator Discriminator looks at both real data and fake data created by the generator. Fake data
73 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator Generator Discriminator looks at both real data and fake data created by the generator. Real data Fake data
74 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
75 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
76 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
77 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
78 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Generator tries to improve its imitation of the data. Real data Fake data
79 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Generator tries to improve its imitation of the data. Real data Fake data
80 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Generator tries to improve its imitation of the data. Real data Fake data
81 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
82 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
83 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
84 6.S191 Introduction to Deep Learning Intuition behind GANs GeneratorDiscriminator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
85 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Generator tries to improve its imitation of the data. Real data Fake data
86 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator ! "#$% = 1 Generator Generator tries to improve its imitation of the data. Real data Fake data
87 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator ! "#$% = 1 Generator Generator tries to improve its imitation of the data. Real data Fake data
88 6.S191 Introduction to Deep Learning Intuition behind GANs Discriminator ! "#$% = 1 Real data Fake data Generator Discriminator tries to identify real data from fakes created by the generator. Generator tries to create imitations of data to trick the discriminator.
89 6.S191 Introduction to Deep Learning Training GANs Discriminator tries to identify real data from fakes created by the generator. Generator tries to create imitations of data to trick the discriminator. min $% max $( )*~,(-.- log 2$( 3 + )5~,(5) log 1 − 2$( :$% (;) Train GAN jointly via minimax game: Discriminator wants to maximize objective s.t. 2 3 close to 1, 2 :(; ) close to 0. Generator wants to minimize objective s.t. 2 :(; ) close to 1.
90 6.S191 Introduction to Deep Learning Why GANs? A. Courville, 6S191 2018.
91 6.S191 Introduction to Deep Learning Why GANs? A. Courville, 6S191 2018.
92 6.S191 Introduction to Deep Learning Generating new data with GANs After training, use generator network to create new data that’s never been seen before. !noise " # $%&'($)'*& +
93 6.S191 Introduction to Deep Learning Deep Generative Modeling: Summary Autoencoders and Variational Autoencoders (VAEs) Generative Adversarial Networks (GANs) Competing generator and discriminator networks Learn lower-dimensional latent space and sample to generate input reconstructions
95 Cool stuffs to do with GANs
CycleGAN Failure Case
97 GANs: Recent Advances
98 6.S191 Introduction to Deep Learning Progressive growing of GANs (NVIDIA) Karras et al., ICLR 2018.
99 6.S191 Introduction to Deep Learning Style-based generator: results Karras et al.,Arxiv 2018.
100 6.S191 Introduction to Deep Learning Style-based transfer: results Karras et al.,Arxiv 2018.
101
102

More Related Content

PPTX
PyData NYC 2019
byLi Jin
 
PDF
Spark AI 2020
byLi Jin
 
PDF
Machine Learning Interpretability / Explainability
byRaouf KESKES
 
PDF
KDD Cup 2021 時系列異常検知コンペ 参加報告
byGentaYoshimura
 
PDF
[系列活動] 手把手的深度學實務
by台灣資料科學年會
 
PDF
building_games_with_ruby_rubyconf
bytutorialsruby
 
PDF
Advanced python
byEU Edge
 
PPT
Cpp tutorial
byFALLEE31188
 
PyData NYC 2019
byLi Jin
 
Spark AI 2020
byLi Jin
 
Machine Learning Interpretability / Explainability
byRaouf KESKES
 
KDD Cup 2021 時系列異常検知コンペ 参加報告
byGentaYoshimura
 
[系列活動] 手把手的深度學實務
by台灣資料科學年會
 
building_games_with_ruby_rubyconf
bytutorialsruby
 
Advanced python
byEU Edge
 
Cpp tutorial
byFALLEE31188
 

Similar to AI 로봇 아티스트의 비밀(창원대학교 정보통신공학과 특강)

PDF
VAE-type Deep Generative Models
byKenta Oono
 
PPTX
Deep learning from a novice perspective
byAnirban Santara
 
PDF
Variational Autoencoders VAE - Santiago Pascual - UPC Barcelona 2018
byUniversitat Politècnica de Catalunya
 
PDF
Deep image generating models
byLuba Elliott
 
PDF
Deep Generative Models I (DLAI D9L2 2017 UPC Deep Learning for Artificial Int...
byUniversitat Politècnica de Catalunya
 
PDF
Introduction to Variational Auto Encoder
byvaidehimadaan041
 
PDF
Introduction to Deep Generative Models
byHao-Wen (Herman) Dong
 
PPTX
GDC2019 - SEED - Towards Deep Generative Models in Game Development
byElectronic Arts / DICE
 
PDF
Alberto Massidda - Scenes from a memory - Codemotion Rome 2019
byCodemotion
 
PPTX
Introduction to Autoencoders: Types and Applications
byAmr Rashed
 
PDF
GEN AI & DL Day-04 & 05.pdf vk kch FF gudiya ki jgh i
bysulabchatterjee475
 
PDF
Lecture 7-8 From Autoencoder to VAE.pdf
byEmadAbdelkader5
 
PDF
Explanation of Autoencoder to Variontal Auto Encoder
byseshathirid
 
PDF
Variational Autoencoders For Image Generation
byJason Anderson
 
PPTX
Piotr Mirowski - Review Autoencoders (Deep Learning) - CIUUK14
byDaniel Lewis
 
PDF
Explicit Density Models
bySangwoo Mo
 
PDF
Deep Generative Models
byChia-Wen Cheng
 
PDF
Tutorial on Theory and Application of Generative Adversarial Networks
byMLReview
 
PPTX
Applications of Variational Autoencoders
byDrEmmanuelJoy2044
 
PPTX
Adversarial Variational Autoencoders to extend and improve generative model
byLoc Nguyen
 
VAE-type Deep Generative Models
byKenta Oono
 
Deep learning from a novice perspective
byAnirban Santara
 
Variational Autoencoders VAE - Santiago Pascual - UPC Barcelona 2018
byUniversitat Politècnica de Catalunya
 
Deep image generating models
byLuba Elliott
 
Deep Generative Models I (DLAI D9L2 2017 UPC Deep Learning for Artificial Int...
byUniversitat Politècnica de Catalunya
 
Introduction to Variational Auto Encoder
byvaidehimadaan041
 
Introduction to Deep Generative Models
byHao-Wen (Herman) Dong
 
GDC2019 - SEED - Towards Deep Generative Models in Game Development
byElectronic Arts / DICE
 
Alberto Massidda - Scenes from a memory - Codemotion Rome 2019
byCodemotion
 
Introduction to Autoencoders: Types and Applications
byAmr Rashed
 
GEN AI & DL Day-04 & 05.pdf vk kch FF gudiya ki jgh i
bysulabchatterjee475
 
Lecture 7-8 From Autoencoder to VAE.pdf
byEmadAbdelkader5
 
Explanation of Autoencoder to Variontal Auto Encoder
byseshathirid
 
Variational Autoencoders For Image Generation
byJason Anderson
 
Piotr Mirowski - Review Autoencoders (Deep Learning) - CIUUK14
byDaniel Lewis
 
Explicit Density Models
bySangwoo Mo
 
Deep Generative Models
byChia-Wen Cheng
 
Tutorial on Theory and Application of Generative Adversarial Networks
byMLReview
 
Applications of Variational Autoencoders
byDrEmmanuelJoy2044
 
Adversarial Variational Autoencoders to extend and improve generative model
byLoc Nguyen
 

More from Changwon National University

PDF
생성인공지능둘러보기.pdf
byChangwon National University
 
PDF
2011 app center Changwon National Univ.
byChangwon National University
 
PDF
인공지능의 파도가 온다
byChangwon National University
 
PDF
Mobile Healthcare Application
byChangwon National University
 
PDF
바다 즐기기
byChangwon National University
 
PDF
알아두면 편리한 macOS 에디터 단축키와 기능
byChangwon National University
 
PDF
키보드 기호의 이름 알아보기(한국어, 영어)
byChangwon National University
 
PDF
20 2 강의를 마치며
byChangwon National University
 
PDF
20 1 코딩스타일
byChangwon National University
 
PDF
18 2 파이썬표준라이브러리
byChangwon National University
 
PDF
18 1 파이썬패키지
byChangwon National University
 
PDF
17 2 필터함수와 맵함수
byChangwon National University
 
PDF
17 1 람다함수
byChangwon National University
 
PDF
16 1 상속과super()
byChangwon National University
 
PDF
15 2 클래스정의와self
byChangwon National University
 
PDF
14 4 슬라이싱
byChangwon National University
 
PDF
14 2 iterator
byChangwon National University
 
PDF
14 3 리스트함수
byChangwon National University
 
PDF
14 1 리스트의 메소드
byChangwon National University
 
PDF
12 3 raise문
byChangwon National University
 
생성인공지능둘러보기.pdf
byChangwon National University
 
2011 app center Changwon National Univ.
byChangwon National University
 
인공지능의 파도가 온다
byChangwon National University
 
Mobile Healthcare Application
byChangwon National University
 
바다 즐기기
byChangwon National University
 
알아두면 편리한 macOS 에디터 단축키와 기능
byChangwon National University
 
키보드 기호의 이름 알아보기(한국어, 영어)
byChangwon National University
 
20 2 강의를 마치며
byChangwon National University
 
20 1 코딩스타일
byChangwon National University
 
18 2 파이썬표준라이브러리
byChangwon National University
 
18 1 파이썬패키지
byChangwon National University
 
17 2 필터함수와 맵함수
byChangwon National University
 
17 1 람다함수
byChangwon National University
 
16 1 상속과super()
byChangwon National University
 
15 2 클래스정의와self
byChangwon National University
 
14 4 슬라이싱
byChangwon National University
 
14 2 iterator
byChangwon National University
 
14 3 리스트함수
byChangwon National University
 
14 1 리스트의 메소드
byChangwon National University
 
12 3 raise문
byChangwon National University
 

Recently uploaded

PDF
The Eternal Citadel Architecture, Sacred Geography, and the Resilience of Som...
byAman Kumar Singh
 
PPTX
Operational Amplifier (op - amp), types of op amp
byVaishaliVaishali14
 
PDF
TR095A120SPC 0.95″ AMOLED Display Module, 120×120 Resolution
bySyluxdisplay
 
PPTX
UNIT 1.3-Structure of Polymers_Material Science-.pptx
byDr. Sandip Thorat
 
PDF
Performance Benchmarking Strategies for AI/HPC/EdgeAI
byDanny Jiang
 
PDF
Java Programming core java and basics concepts
byHarish Khodke
 
PDF
Life Cycle Analysis of Electric and Internal Combustion Engine Vehicles
byUniversity Institute of Technology, Barkatullah University
 
PPTX
Diagram of Control Valves used in Industrial Fluid Power.pptx
bySBM Polytechnic
 
PDF
BCS402 MICROCONTROLLERS MODULE 1
byAbhijith L Kotian
 
PDF
22PEOIT4C_Artificial_Intelligence_UNIT_III_Notes_Final
byGuru Nanak Technical Institutions
 
PDF
22PEOIT4C Artificial Intelligence Unit II notes Final.pdf
byGuru Nanak Technical Institutions
 
PPTX
overview of hoists in construction related to architecture
byptripathi2250
 
PDF
Introduction : Operating-System Services
bySanjay Gunjal
 
PDF
TU/e - Lecture Geotechnics - Case Singelgrachtgarage-Marnix - Arkesteijn
byRuud Arkesteijn
 
PPTX
Ion exchange or demineralization to soften water.pptx
byChemical Engineering Dept. NIT Rourkela-769008, Odisha, India
 
PPTX
Symmetrical faults in power system .pptx
byRased Khan
 
PDF
TU/e - 2025-2026 - Lecture Geotechnics 3 - Arkesteijn
byRuud Arkesteijn
 
PPTX
MET 464- MICRO AND AND NANO MANUFACTURING -MOD 2.pptx KTU 2019
byVINAY B
 
PPTX
Design and Development of Friction Stir Welding joints of Aluminium and Coppe...
byNareshKumarSingh12
 
PDF
22PE0IT4C_Artificial_Intelligence_Unit_1_QB_Final
byGuru Nanak Technical Institutions
 
The Eternal Citadel Architecture, Sacred Geography, and the Resilience of Som...
byAman Kumar Singh
 
Operational Amplifier (op - amp), types of op amp
byVaishaliVaishali14
 
TR095A120SPC 0.95″ AMOLED Display Module, 120×120 Resolution
bySyluxdisplay
 
UNIT 1.3-Structure of Polymers_Material Science-.pptx
byDr. Sandip Thorat
 
Performance Benchmarking Strategies for AI/HPC/EdgeAI
byDanny Jiang
 
Java Programming core java and basics concepts
byHarish Khodke
 
Life Cycle Analysis of Electric and Internal Combustion Engine Vehicles
byUniversity Institute of Technology, Barkatullah University
 
Diagram of Control Valves used in Industrial Fluid Power.pptx
bySBM Polytechnic
 
BCS402 MICROCONTROLLERS MODULE 1
byAbhijith L Kotian
 
22PEOIT4C_Artificial_Intelligence_UNIT_III_Notes_Final
byGuru Nanak Technical Institutions
 
22PEOIT4C Artificial Intelligence Unit II notes Final.pdf
byGuru Nanak Technical Institutions
 
overview of hoists in construction related to architecture
byptripathi2250
 
Introduction : Operating-System Services
bySanjay Gunjal
 
TU/e - Lecture Geotechnics - Case Singelgrachtgarage-Marnix - Arkesteijn
byRuud Arkesteijn
 
Ion exchange or demineralization to soften water.pptx
byChemical Engineering Dept. NIT Rourkela-769008, Odisha, India
 
Symmetrical faults in power system .pptx
byRased Khan
 
TU/e - 2025-2026 - Lecture Geotechnics 3 - Arkesteijn
byRuud Arkesteijn
 
MET 464- MICRO AND AND NANO MANUFACTURING -MOD 2.pptx KTU 2019
byVINAY B
 
Design and Development of Friction Stir Welding joints of Aluminium and Coppe...
byNareshKumarSingh12
 
22PE0IT4C_Artificial_Intelligence_Unit_1_QB_Final
byGuru Nanak Technical Institutions
 

AI 로봇 아티스트의 비밀(창원대학교 정보통신공학과 특강)

  • 1.
    AI 로봇 아티스트의비밀 Deep Generative Models 2019-05-23 Sung-Soo Kim sungsoo@etri.re.kr Smart Data Research Group 특강
  • 2.
    Outline ! Introduction – ArtificialIntelligence,MachineLearning,DeepLearning !Autoencoders ! Variational Autoencoders (VAEs) ! Generative Adversarial Networks (GANs) ! Summary
  • 3.
  • 4.
    4 6.S191 Introduction toDeep Learning Which face is fake?
  • 5.
  • 6.
  • 7.
  • 8.
  • 9.
    Dog Cat 입력이 주어지면 출력을내보낸다. 인공지능
  • 10.
    입력이 주어지면, 출력을내보낸다. 우리 이런 걸 어디서 봤죠?
  • 11.
  • 12.
    인공지능 별 거아니죠? 무서울 것 없습니다. 근데 왜 갑자기 무서워 졌을까요?
  • 13.
    모두의연구소 수학적으로 표현할 수없었던 복잡한 인간의 두뇌를 데이터를 기반으로 흉내내다.
  • 14.
  • 15.
  • 16.
    X[n] F[n] F[n-1] Y[n] n :data index Data가 늘어날수록 점점 인공지능 알고리즘이 학습(Learning)한다.
  • 17.
  • 18.
    3 X ??? 6 DogCat Supervised Learning - 정답이 주어진다. - (비교적)문제풀이가 쉽다
  • 19.
    3 X ??? ??? Unsupervised Learning 미지수 2개,방정식 1개 풀 수 있나요?
  • 20.
    Reinforcement Learning - (정답이 아닌)reward가 주어진다. Reward
  • 21.
  • 22.
  • 23.
    뉴런과 인공뉴런 x 1 x 1 b a 다른 머신러닝 기법들과의차이점 1: Nonlinear(복잡한) activation function b a
  • 24.
    다층 레이어 (Multiple Layer) 다른머신러닝 기법들과의 차이점 2 : Nonlinear function의 Nonlinear function의 Nonlinear function … 엄청 복잡한 함수 (인공지능)을 만들 수 있다. Hidden layer가 2개이상인 NN를 Deep Learning이라고 부른다.
  • 25.
  • 26.
    26 6.S191 Introduction toDeep Learning Supervised vs unsupervised learning Supervised Learning Data: (", $) " is data, $ is label Goal: Learn function to map " → $ Examples: Classification, regression, object detection, semantic segmentation, etc. Unsupervised Learning Data: " " is data, no labels! Goal: Learn some hidden or underlying structure of the data Examples: Clustering, feature or dimensionality reduction, etc.
  • 27.
    27 6.S191 Introduction toDeep Learning Supervised vs unsupervised learning Supervised Learning Data: (", $) " is data, $ is label Goal: Learn function to map " → $ Examples: Classification, regression, object detection, semantic segmentation, etc. Unsupervised Learning Data: " " is data, no labels! Goal: Learn some hidden or underlying structure of the data Examples: Clustering, feature or dimensionality reduction, etc.
  • 28.
    28 6.S191 Introduction toDeep Learning Generative modeling Goal: Take as input training samples from some distribution and learn a model that represents that distribution Density Estimation Sample Generation Input samples Generated samples Training data ~ "#$%$ & Generated ~ "'(#)* & How can we learn "'(#)* & similar to "#$%$ & ? samples KeyTakeaway#1 KeyTakeaway#2 MainTakeaways
  • 29.
    29 6.S191 Introduction toDeep Learning Why generative models? Debiasing vs Capable of uncovering underlying latent variables in a dataset Homogeneous skin color, pose Diverse skin color, pose, illumination How can we use latent distributions to create fair and representative datasets?
  • 30.
    30 6.S191 Introduction toDeep Learning Why generative models? Outlier detection 95% of Driving Data: (1) sunny, (2) highway, (3) straight road Detect outliers to avoid unpredictable behavior when training Edge Cases Harsh Weather Pedestrians • Problem: How can we detect when we encounter something new or rare? • Strategy: Leverage generative models, detect outliers in the distribution • Use outliers during training to improve even more!
  • 31.
    31 Latent variable models AutoencodersandVariational Autoencoders (VAEs) Generative Adversarial Networks (GANs)
  • 32.
    32 6.S191 Introduction toDeep Learning What is a latent variable? Myth of the Cave 515c~d
  • 33.
  • 34.
    34 6.S191 Introduction toDeep Learning What is a latent variable? Can we learn the true explanatory factors, e.g. latent variables, from only observed data?
  • 35.
  • 36.
    36 6.S191 Introduction toDeep Learning Autoencoders: background Unsupervised approach for learning a lower-dimensional feature representation from unlabeled training data ! " “Encoder” learns mapping from the data, !, to a low-dimensional latent space, " Why do we care about a low-dimensional "?
  • 37.
    37 6.S191 Introduction toDeep Learning Autoencoders: background How can we learn this latent space? Train the model to use these features to reconstruct the original data ! " “Decoder” learns mapping back from latent, ", to a reconstructed observation, #! #!
  • 38.
    38 6.S191 Introduction toDeep Learning Autoencoders: background How can we learn this latent space? Train the model to use these features to reconstruct the original data ! " #! ℒ !, #! = ! − #! ( Loss function doesn’t use any labels!!
  • 39.
    39 6.S191 Introduction toDeep Learning Autoencoders: background How can we learn this latent space? Train the model to use these features to reconstruct the original data ! " #! ℒ !, #! = ! − #! ( Loss function doesn’t use any labels!!
  • 40.
    40 6.S191 Introduction toDeep Learning Dimensionality of latent space à reconstruction quality 2D latent space 5D latent space GroundTruth Autoencoding is a form of compression! Smaller latent space will force a larger training bottleneck
  • 41.
    41 6.S191 Introduction toDeep Learning Autoencoders for representation learning Bottleneck hidden layer forces network to learn a compressed latent representation Reconstruction loss forces the latent representation to capture (or encode) as much “information” about the data as possible Autoencoding = Automatically encoding data
  • 42.
  • 43.
    43 2019 ModuLabs Taxonomy of GenerativeModels Figure credit: Ian Goodfellow, NIPS 2016 Tutorial: Generative Adversarial Networks
  • 44.
    44 6.S191 Introduction toDeep Learning VAEs: key difference with traditional autoencoder ! " #!
  • 45.
    45 6.S191 Introduction toDeep Learning VAEs: key difference with traditional autoencoder ! " #! $ %
  • 46.
    46 6.S191 Introduction toDeep Learning VAEs: key difference with traditional autoencoder ! " #! $ % mean vector standard deviation vector Variational autoencoders are a probabilistic twist on autoencoders! Sample from the mean and standard dev. to compute latent sample
  • 47.
    47 6.S191 Introduction toDeep Learning VAE optimization ! " #! $ % Encoder computes: &'(z|!) Decoder computes: ,-(x|")
  • 48.
    48 6.S191 Introduction toDeep Learning VAE optimization ! " #! $ % Encoder computes: &'(z|!) Decoder computes:,-(x|") ℒ ϕ, 2 = (reconstruction loss) + (regularization term)
  • 49.
    49 6.S191 Introduction toDeep Learning VAE optimization ! " #! $ % Encoder computes: &'(z|!) Decoder computes:,-(x|") ℒ ϕ, 2, ! = (reconstruction loss) + (regularization term)
  • 50.
    50 6.S191 Introduction toDeep Learning VAE optimization ! " #! $ % Encoder computes: &'(z|!) Decoder computes:,-(x|") ℒ ϕ, 2, ! = (reconstruction loss) + (regularization term) e.g. ! − #! 5
  • 51.
    51 6.S191 Introduction toDeep Learning VAE optimization ! " #! $ % Encoder computes: &'(z|!) Decoder computes:,-(x|") ℒ ϕ, 2, ! = (reconstruction loss) + (regularization term) 4 &' z ! ∥ & " Inferred latent distribution Fixed prior on latent distribution
  • 52.
    52 6.S191 Introduction toDeep Learning Priors on the latent distribution ! "# z % ∥ " ' Inferred latent distribution Fixed prior on latent distribution Common choice of prior: " ' = ) * = 0, -. = 1 • Encourages encodings to distribute encodings evenly around the center of the latent space • Penalize the network when it tries to “cheat” by clustering points in specific regions (ie. memorizing the data)
  • 53.
    53 6.S191 Introduction toDeep Learning Priors on the latent distribution ! "# z % ∥ " ' Common choice of prior: " ' = ) * = 0, -. = 1 • Encourages encodings to distribute encodings evenly around the center of the latent space • Penalize the network when it tries to “cheat” by clustering points in specific regions (ie. memorizing the data) = − 1 2 2 345 678 -3 + *3 . − 1 − log -3 KL-divergence between the two distributions
  • 54.
    54 6.S191 Introduction toDeep Learning VAEs computation graph ! " #! $ % Encoder computes: &'(z|!) Decoder computes:,-(x|") ℒ ϕ, 2, ! = (reconstruction loss) + (regularization term)
  • 55.
    55 6.S191 Introduction toDeep Learning VAEs computation graph ! " #! $ % Encoder computes: &'(z|!) Decoder computes:,-(x|") ℒ ϕ, 2, ! = (reconstruction loss) + (regularization term) Problem: We cannot backpropagate gradients through sampling layers!
  • 56.
    56 6.S191 Introduction toDeep Learning Reparametrizing the sampling layer ! " # Key Idea: ! ~%(", #() Consider the sampled latent vector as a sum of • a fixed " vector, • and fixed # vector, scaled by random constants drawn from the prior distribution ⇒ ! = " + #⨀. where .~%(0,1)
  • 57.
    57 6.S191 Introduction toDeep Learning Reparametrizing the sampling layer ! "" ∼ $%(z|)) + ) Deterministic node Stochastic node Original form Backprop
  • 58.
    58 6.S191 Introduction toDeep Learning Reparametrizing the sampling layer ! "" ∼ $%(z|)) ! " " = ,(-, ), /) - ) /- ) Deterministic node Stochastic node Original form Reparametrized form 0! 0" 0! 0- Backprop ~2(0,1)
  • 59.
    59 6.S191 Introduction toDeep Learning VAEs: Latent perturbation Slowly increase or decrease a single latent variable Keep all other variables fixed Head pose Different dimensions of ! encodes different interpretable latent features
  • 60.
    60 6.S191 Introduction toDeep Learning VAEs: Latent perturbation Head pose Smile Ideally, we want latent variables that are uncorrelated with each other Enforce diagonal prior on the latent variables to encourage independence Disentanglement
  • 61.
    61 6.S191 Introduction toDeep Learning VAEs: Latent perturbation Google BeatBlender GoogleBeatBlender
  • 62.
  • 63.
    63 6.S191 Introduction toDeep Learning VAE summary 1. Compress representation of world to something we can use to learn 2. Reconstruction allows for unsupervised learning (no labels!) 3. Reparameterization trick to train end-to-end 4. Interpret hidden latent variables using perturbation 5. Generating new examples ! " #!
  • 64.
  • 65.
    65 2019 ModuLabs “The mostimportant idea in Machine Learning in the last 20 years”
  • 66.
    66 2019 ModuLabs Taxonomy of GenerativeModels Figure credit: Ian Goodfellow, NIPS 2016 Tutorial: Generative Adversarial Networks
  • 67.
    67 6.S191 Introduction toDeep Learning What if we just want to sample? Idea: don’t explicitly model density, and instead just sample to generate new instances. Problem: want to sample from complex distribution – can’t do this directly! Solution: sample from something simple (noise), learn a transformation to the training distribution. !noise " #$%&' “fake” sample from the training distribution Generator Network "
  • 68.
    68 6.S191 Introduction toDeep Learning Generative Adversarial Networks (GANs) Generative Adversarial Networks (GANs) are a way to make a generative model by having two neural networks compete with each other. The discriminator tries to identify real data from fakes created by the generator. The generator turns noise into an imitation of the data to try to trick the discriminator. !noise " # $%&'($)'*& +
  • 69.
  • 70.
  • 71.
    71 6.S191 Introduction toDeep Learning Intuition behind GANs Generator Generator starts from noise to try to create an imitation of the data. Fake data OneMoreTime!
  • 72.
    72 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator Generator Discriminator looks at both real data and fake data created by the generator. Fake data
  • 73.
    73 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator Generator Discriminator looks at both real data and fake data created by the generator. Real data Fake data
  • 74.
    74 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
  • 75.
    75 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
  • 76.
    76 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
  • 77.
    77 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
  • 78.
    78 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Generator tries to improve its imitation of the data. Real data Fake data
  • 79.
    79 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Generator tries to improve its imitation of the data. Real data Fake data
  • 80.
    80 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Generator tries to improve its imitation of the data. Real data Fake data
  • 81.
    81 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
  • 82.
    82 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
  • 83.
    83 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
  • 84.
    84 6.S191 Introduction toDeep Learning Intuition behind GANs GeneratorDiscriminator ! "#$% = 1 Discriminator tries to predict what’s real and what’s fake. Real data Fake data
  • 85.
    85 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator Generator ! "#$% = 1 Generator tries to improve its imitation of the data. Real data Fake data
  • 86.
    86 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator ! "#$% = 1 Generator Generator tries to improve its imitation of the data. Real data Fake data
  • 87.
    87 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator ! "#$% = 1 Generator Generator tries to improve its imitation of the data. Real data Fake data
  • 88.
    88 6.S191 Introduction toDeep Learning Intuition behind GANs Discriminator ! "#$% = 1 Real data Fake data Generator Discriminator tries to identify real data from fakes created by the generator. Generator tries to create imitations of data to trick the discriminator.
  • 89.
    89 6.S191 Introduction toDeep Learning Training GANs Discriminator tries to identify real data from fakes created by the generator. Generator tries to create imitations of data to trick the discriminator. min $% max $( )*~,(-.- log 2$( 3 + )5~,(5) log 1 − 2$( :$% (;) Train GAN jointly via minimax game: Discriminator wants to maximize objective s.t. 2 3 close to 1, 2 :(; ) close to 0. Generator wants to minimize objective s.t. 2 :(; ) close to 1.
  • 90.
    90 6.S191 Introduction toDeep Learning Why GANs? A. Courville, 6S191 2018.
  • 91.
    91 6.S191 Introduction toDeep Learning Why GANs? A. Courville, 6S191 2018.
  • 92.
    92 6.S191 Introduction toDeep Learning Generating new data with GANs After training, use generator network to create new data that’s never been seen before. !noise " # $%&'($)'*& +
  • 93.
    93 6.S191 Introduction toDeep Learning Deep Generative Modeling: Summary Autoencoders and Variational Autoencoders (VAEs) Generative Adversarial Networks (GANs) Competing generator and discriminator networks Learn lower-dimensional latent space and sample to generate input reconstructions
  • 95.
    95 Cool stuffs todo with GANs
  • 96.
  • 97.
  • 98.
    98 6.S191 Introduction toDeep Learning Progressive growing of GANs (NVIDIA) Karras et al., ICLR 2018.
  • 99.
    99 6.S191 Introduction toDeep Learning Style-based generator: results Karras et al.,Arxiv 2018.
  • 100.
    100 6.S191 Introduction toDeep Learning Style-based transfer: results Karras et al.,Arxiv 2018.
  • 101.
  • 102.