NAVER Engineering, profile picture
Uploaded byNAVER Engineering
PDF, PPTX8,154 views

그림 그리는 AI

The document covers advancements in image recognition and generation technologies, including facial and object recognition, using deep learning and big data techniques since 2015. It discusses various models like GANs (Generative Adversarial Networks) and VAEs (Variational Autoencoders) that have emerged, focusing on their applications in generating realistic images from textual descriptions. The content also highlights comparative analyses of these models, their learning methods, and their performance in generating high-quality visuals.

Embed presentation

Download as PDF, PPTX
NAVER Clova 이활석
인식 이미지 내의 존재하는 정보를 찾는 기술 얼굴 인식 홍채 인식 번호판 인식 지문 인식 생성 특정 정보를 담는 이미지를 생성하는 기술 스타일 변환스타일 이미지 컴퓨터 비전 기술INTRODUCTION
이미지 인식 Image Recognition
ü ImageNet 대회 • 1000 종류의 클래스 존재 u 목표 : 이미지 내 존재하는 객체 분류 < 사람=5% u 2015년부터 사람보다 잘 하기 시작 (2017년으로 종료) 빅데이터+딥러닝 > 사람 1 / 6 RECOGNITION 대회 현황IMAGENET 사람이 5% 오차? : 정답의 오류 및 종류의 세분화
정해진 수의 물건들의 특징을 잘 파악하여 구분할 수 있다! 딥뉴럴 네트워크 개와 고양이를 구분하는 모델 RECOGNITION 2015년부터 AI가 잘 할 수 있게 된 것은?IMAGENET 2 / 6
RECOGNITION 이미지 태그 생성APPLICATION 이미지와 관련 있는 태그 찾기 https://www.clarifai.com/demo 3 / 6
RECOGNITION 객체 분류APPLICATION ü 사각형 영역 단위로 구분 (RCNN, ‘15.04) ü 화소 단위로 구분 (DeepLab, ‘17.03) 4 / 6
RECOGNITION 신체 부위 분류APPLICATION ü Openpose, ‘17.04 오승환 투구 슬로우 모션 영상 https://github.com/CMU-Perceptual-Computing-Lab/openpose 5 / 6 얼굴 + 신체 + 손
이미지 생성 Image Generation
딥뉴럴 네트워크 개와 고양이를 생성할 수 있는 모델 분류 모델 보다는 개와 고양이를 제대로 이해하고 있다 Generative Model “What I cannot create, I do not understand.” - Richard Feynman 1 / 13 GENERATION 2015년부터 AI 연구자들이 관심 갖게 된 것은?INTORUDCTION
StackGAN : Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks , ‘16.12 AI가 생성한 사진을 찾아보세요! (1) This flower has overlapping pink pointed petals surrounding a ring of short yellow filaments 이 꽃은 짧은 노란색 필라멘트의 고리를 둘러싼 핑크색 뾰족한 꽃잎이 겹쳐져 있습니다 (2) This flower has upturned petals which are thin and orange with rounded edges 이 꽃은 둥근 모서리를 가진 얇고 오렌지색의 꽃잎이 위로 향해 있습니다 (3) A flower with small pink petals and a massive central orange and black stamen cluster 작은 분홍색 꽃잎들과 중심에 다수의 오렌지색과 검은 색 수술 군집이 있는 꽃 (1) (2) (3) 2 / 13 GENERATION Quiz : stackGANINTORUDCTION
3 / 13 GENERATION Generative ModelsINTORUDCTION 𝑧 Generative Model 0.1 0.7 -2 … 1.4 𝑥 𝑧 Discriminative Model 0.1 0.3 0 … 0.4 𝑥 비교사학습 Un-supervised Learning 교사학습 Supervised Learning 고양이 강아지 여우 뱀 Latent Vector Noise Code Feature Data (Image) 구분 모델 생성 모델 Supervised VS Un-supervised
4 / 13 GENERATION Generative ModelsINTORUDCTION Density Estimation https://github.com/mingyuliutw/cvpr2017_gan_tutorial/blob/master/gan_tutorial.pdf Training Examples Maximum Likelihood Sampling 𝑝(𝑥)
5 / 13 GENERATION Generative ModelsINTORUDCTION Taxomomy http://www.iangoodfellow.com/slides/2016-12-04-NIPS.pdf 분포에 대한 모델을 가정하고 그 모델을 결정하는 파라미터를 조절해 가며 타겟 분포가 되도록 노력 생성되는 샘플들의 분포가 타겟 분포가 되도록 노력 VAE GAN
VAE Varaitional AutoEncoder
Variational • 𝒑(𝒙)를 구할 때 variational learning/inference 활용 AutoEncoder 1 / 33 GENERATION 용어의 의미VAE Auto-Encoding Variational Bayes , ‘14.03 Latent Variable Target Data 𝑥 𝑔)* 𝑧 𝑔)+ 𝑧 𝑝 𝑥 𝑔)+ 𝑧 < 𝑝 𝑥 𝑔)* 𝑧 확률 분포를 정해주는 파라미터를 추정. (예. 가우시안일 경우 평균,표준편차) 𝑧 𝑔)(𝑧)Generator gθ(.) 𝑝 𝑧 𝑝 𝑥 𝑔)(𝑧) sampling modeling - 𝑝 𝑥 𝑔)(𝑧) 𝑝 𝑧 𝑑𝑧 = 𝑝(𝑥) variational learning - 𝑝 𝑥 𝑔)(𝑧) 𝑝 𝑧 𝑑𝑧 = 𝐸2(3)[𝑝)(𝑥|𝑧)]
Variational • 𝒑(𝒙)를 구할 때 variational learning/inference 활용 AutoEncoder 2 / 33 GENERATION 용어의 의미VAE Auto-Encoding Variational Bayes , ‘14.03 𝑧 𝑔)(𝑧)Generator gθ(.) 𝑝 𝑧 sampling • 입력과 출력이 동일한 네트워크 𝑝 𝑧|𝑥 x를 잘 생성해는 z가 샘플링 잘 되도록, x를 힌트로 제공함 𝑝 𝑧|𝑞8(𝑥) p(z|x)의 모델을 가정하고, 그 모델을 결정짓는 파라미터를 네트워크로 추정 Decoder gθ(.) Encoder qφ(.) 𝑥 𝑝 𝑧|𝑞8(𝑥) 𝑞8(𝑥) sampling 𝑧 𝑔) 𝑧 𝑥→
3 / 33 GENERATION Variational InferenceVAE Variational Lower Bound (Evidence Lower BOund, ELBO) log 𝑝 𝑥 = 𝐸𝐿𝐵𝑂 + 𝐾𝐿(𝑞8 𝑧|𝑥 ∥ 𝑝 𝑧|𝑥 ) log 𝑝 𝑥 𝐾𝐿(𝑞8 𝑧|𝑥 ∥ 𝑝 𝑧|𝑥 ) 𝐸𝐿𝐵𝑂(𝜙) 𝜙D 𝜙E Optimization Problem 1 on 𝜙: Variational Inference argmin 8 𝐾𝐿(𝑞8 𝑧|𝑥 ∥ 𝑝 𝑧|𝑥 ) == argmax 8 𝐸𝐿𝐵𝑂(𝜙) log 𝑝 𝑥 ≥ 𝔼NO 3|P log 𝑝) 𝑥|𝑧 − 𝐾𝐿(𝑞8 𝑧|𝑥 ∥ 𝑝 𝑧 ) = 𝐸𝐿𝐵𝑂 Optimization Problem 2 on 𝜃: Variational Learning argmax ) 𝔼NO 3|P log 𝑝) 𝑥|𝑧 = argmax ) 𝐸𝐿𝐵𝑂(𝜃) Final Optimization Problem argmax ),8 𝐸𝐿𝐵𝑂(𝜃, 𝜙) • KL : Kullback-Leibler Divergence 두 확률 분포가 얼마나 다른지에 대한 척도
𝐿U 𝜙, 𝜃, 𝑥U = −𝔼NO 3|P log 𝑝) 𝑥U|𝑧 + 𝐾𝐿(𝑞8 𝑧|𝑥U ∥ 𝑝 𝑧 ) Reconstruction Error 복원 오차 입력과 출력 간의 cross-entropy Regularization 제약 조건 Prior분포와의 다른 정도 [ VAE의 특징들 ] 1. Decoder가 최소한 학습 데이터는 생성해 낼 수 있게 된다. à 생성된 데이터가 학습 데이터 좀 닮아 있다. 2. Encoder가 최소한 학습 데이터는 잘 latent vector로 표현할 수 있게 된다. à 데이터의 추상화를 위해 많이 사용된다. 4 / 33 GENERATION Encoder & DecoderVAE 𝑥U Encoder 𝑞8(𝑧|𝑥U) 𝜇U 𝜎U 𝜖~𝑁(0, 𝐼) 𝜖U 𝑧U Decoder 𝑝) 𝑥|𝑧 𝑦U = 𝑥U Bernoulli Gaussian Cost Function
28 𝑥U 𝑝U 𝐷 28 𝐷=784 MLP with 2 hidden layers (500, 500) Architecture 5 / 33 GENERATION Result : MNISTVAE 𝑔)(_)𝑞8 _ 𝑞8 𝑧|𝑥 ~𝑧 𝑔) 𝑥|𝑧 Encoder Posterior Inference Network Decoder Generator Generation Network SAMPLING 𝜇U 𝜎U Latent Space
Reproduce Input image |z| =2 |z| =5 |z| =20 https://github.com/hwalsuklee/tensorflow-mnist-VAE 6 / 33 GENERATION Result : MNISTVAE
Denoising Input image + zero-masking noise with 50% prob. + salt&peppr noise with 50% prob. Restored image https://github.com/hwalsuklee/tensorflow-mnist-VAE 7 / 33 GENERATION Result : MNISTVAE
Learned Manifold 학습이 잘 되었을 수록 2D공간에서 같은 숫자들을 생성하는 z들은 뭉쳐있고, 다른 숫자들은 생성하는 z들은 떨어져 있어야 한다. https://github.com/hwalsuklee/tensorflow-mnist-VAE 8 / 33 GENERATION Result : MNISTVAE Decoding / GenerationEncoding / Feature Extraction
9 / 33 GENERATION DEMO : MNIST / Gray FaceVAE http://www.dpkingma.com/sgvb_mnist_demo/demo.html |z|=12 24 24 Handwritten Digits Generation http://vdumoulin.github.io/morphing_faces/online_demo.html |z|=29 64 64 Gray Face Generation
https://magenta.tensorflow.org/sketch-rnn-demo https://magenta.tensorflow.org/assets/sketch_rnn_demo/multi_vae.html 10 / 33 GENERATION DEMO : Sketch RNNVAE
IntroductionCVAE Conditional VAE 11 / 33 GENERATION
input CVAE, epoch 1 VAE, epoch 1 CVAE, epoch 20 VAE, epoch 20 Reproduce |z| = 2 MNIST resultsCVAE https://github.com/hwalsuklee/tensorflow-mnist-CVAE 12 / 33 GENERATION
CVAE, epoch 1 VAE, epoch 1 CVAE, epoch 20 VAE, epoch 20 input Denoising |z| = 2 MNIST resultsCVAE https://github.com/hwalsuklee/tensorflow-mnist-CVAE 13 / 33 GENERATION
Handwriting styles obtained by fixing the class label and varying z |z| = 2 y=[1,0,0,0,0,0,0,0,0,0] y=[0,1,0,0,0,0,0,0,0,0] y=[0,0,1,0,0,0,0,0,0,0] y=[0,0,0,1,0,0,0,0,0,0] y=[0,0,0,0,1,0,0,0,0,0] y=[0,0,0,0,0,0,1,0,0,0]y=[0,0,0,0,0,1,0,0,0,0] y=[0,0,0,0,0,0,0,0,1,0]y=[0,0,0,0,0,0,0,1,0,0] y=[0,0,0,0,0,0,0,0,0,1] MNIST resultsCVAE https://github.com/hwalsuklee/tensorflow-mnist-CVAE 14 / 33 GENERATION
Z-sampling 각 행 별로, 고정된 z값에 대해서 label정보만 바꿔 서 이미지 생성 (스타일 유지하면 숫자만 바뀜) MNIST resultsCVAE Analogies : Result in paper Semi-Supervised Learning with Deep Generative Models : https://arxiv.org/abs/1406.5298 15 / 33 GENERATION
𝑧D 𝑧E 𝑧` 𝑧a 𝑐c 𝑐D 𝑐E 𝑐` 𝑐a 𝑐d 𝑐e 𝑐f 𝑐g 𝑐h Handwriting style for a given z must be preserved for all labels Analogies |z| = 2 MNIST resultsCVAE https://github.com/hwalsuklee/tensorflow-mnist-CVAE 𝑐c 𝑐D 𝑐E 𝑐` 𝑐a 𝑐d 𝑐e 𝑐f 𝑐g 𝑐h Real handwritten image 실제로 손으로 쓴 글씨 ‘3’을 CVAE의 label정보와 같이 넣었을 때 얻는 latent vector는 decoder의 고정 입력으로 하고, label정보만 바꿨을 경우 16 / 33 GENERATION
GAN Generative Adversarial Networks
17 / 33 GENERATION 용어의 의미GAN Generative Adversarial Networks , ‘14.06 Generative • Generator (생성자) 및 Discriminator가 (구분자) 서로 적대적으로 행동 • 생성 모델 Adversarial Networks • 생성자 및 구분자 모두 딥뉴럴 네트워크로 모델링 x 0~1D G G(z) Discriminator Generator z “The coolest idea in ML in the last twenty years” – Yann LeCun
생성자 Generator 구별자 Discriminator G A Networks 두 개의 네트워크 dversarial 적대적으로 학습된 enerative 데이터 생성목적으로 실제 / 가짜 가짜 사진 실제 데이터 실제 사진 구별자는 생성자의 결과를 거짓이라고 판단하려고 노력 생성자는 구별자가 참이라고 판단하게 하려고 노력 거짓말 쟁이 거짓말 탐지기 적대적 18 / 33 GENERATION 꽃 사진 생성 예시GAN
Noise (Latent variable) 생성자 Generator 구별자 Discriminator 실 데이터 생성자 Yes / No 실 데이터? Generator가 매 번 다른 샘플을 만들어 내기 위한 입력 정답 얼굴 DB 구별자 성능 향상 생성자 성능 향상 19 / 33 GENERATION 얼굴 사진 생성 예시GAN
Noise (Latent variable) eneratorDiscriminator Data sample Yes / No G 𝑝ijkj(𝑥) 𝐷 𝑥 = 1 𝐷 𝐺 𝑧 = 0 𝑥 𝐺(𝑧) 𝑧 𝐷 𝐺 𝑧 =1 𝑝3(𝑧) 𝑉 𝐷, 𝐺 = 𝔼P~2opqp(P) log𝐷(𝑥) + 𝔼3~2r(3) log 1 − 𝐷 𝐺 𝑧Value function of GAN Goal 𝐷∗ , 𝐺∗ = min t max u 𝑉 𝐷, 𝐺 minmax problem!!! 20 / 33 GENERATION ProblemGAN Notation
D max u 𝑉 𝐷, 𝐺 = max u 𝔼P~2opqp log𝐷(𝑥) + 𝔼3~2r log 1 − 𝐷 𝐺 𝑧 G is fixed Maximize prob. of D(real) Minimize prob. of D(fake) G min t 𝑉 𝐷, 𝐺 = min t 𝔼3~2r log 1 − 𝐷 𝐺 𝑧 D is fixed Maximize prob. of D(fake) = max t 𝔼3~2r log 𝐷 𝐺 𝑧 Alternating Optimization 21 / 33 GENERATION ProblemGAN Optimization 𝑉 𝐷, 𝐺 = 𝔼P~2opqp(P) log𝐷(𝑥) + 𝔼3~2r(3) log 1 − 𝐷 𝐺 𝑧Value function of GAN Goal 𝐷∗ , 𝐺∗ = min t max u 𝑉 𝐷, 𝐺
𝐷∗ 𝑥 = 𝑝ijkj(𝑥) 𝑝ijkj 𝑥 + 𝑝v(𝑥)D 𝑝v ∗ 𝑥 = 𝑝ijkj(𝑥) G D(x) p_data p_g(x) Optimization for D(x) Optimization for G(z) Equilibrium : D(x)=0.5 G(z) maps z to x 1D Gaussian Approximation example 22 / 33 GENERATION ProblemGAN Optimal Solution http://cs.stanford.edu/people/karpathy/gan/ Uniform Distribution https://github.com/hwalsuklee/tensorflow-GAN-1d-gaussian-ex
23 / 33 GENERATION VAE VS GANGAN Model Optimization Image Quality Generalization VAE • Stochastic gradient descent • Converge to local minimum • Easier • Smooth • Blurry • Tend to remember input images GAN • Alternating stochastic gradient descent • Converge to saddle points • Harder v Model collapsing v Unstable convergence • Sharp • Artifact • Generate new unseen images http://aliensunmin.github.io/project/accv16tutorial/media/generative.pdf z DEx x DecoderEncoder x 0~1D G G(z) Discriminator Generator z VAE GAN
Comparison between VAE vs GAN VAE : maximum likelihood approach GAN http://videolectures.net/site/normal_dl/tag=1129740/deeplearning2017_courville_generative_models_01.pdf 24 / 33 GENERATION VAE VS GANGAN
이후로 소개되는 모든 논문들은 DCGAN에서 크게 벗어나지 않음 대부분의 상황에서 안정적으로 학습이 되는 Deep Convolutional GAN구조를 제안 Generator Discriminator Pooling Layers Not used. But use strided convolutions instead. Batch normalization Use except output layer Use except input layer Fully connected hidden layers Not used Activation function ReLU for all layers except for the output, which uses Tanh LeakyReLU for all layers 25 / 33 GENERATION Key ContributionDCGAN Unsupervised Representation learning with Deep Convolutional Generative Adversarial Networks (DCGAN), ‘15.11
26 / 33 GENERATION 네트워크 구조DCGAN 생성자 Generator 구별자 Discriminator
27 / 33 GENERATION Face Generation ResultsDCGAN http://carpedm20.github.io/faces/ 유명인 얼굴 DB LEVEL 2 LEVEL 1 진짜 진짜
28 / 33 GENERATION Face Generation Results : InterpolationDCGAN z1 z2(z1, z2)를 보간 후 새로 생성한 z에 대한 결과
z1 : 썬글라스를 낀 남자들을 발생시키는 z값 평균 z2 : 썬글라스를 안 낀 남자들을 발생시키는 z값 평균 z3 : 썬글라스를 안 낀 여자들을 발생시키는 z값 평균 z1 z2 z3 z1-z2+z3 썬글라스를 낀 여자들 29 / 33 GENERATION Face Generation Results : Vector arithmeticDCGAN
30 / 33 GENERATION Character Generation ResultsDCGAN http://mattya.github.io/chainer-DCGAN 만화 캐릭터 DB
31 / 33 GENERATION A List of Some PapersGAN zoo https://deephunt.in/the-gan-zoo-79597dc8c347 GAN — Generative Adversarial Networks 3D-GAN — Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling acGAN — Face Aging With Conditional Generative Adversarial Networks AC-GAN — Conditional Image Synthesis With Auxiliary Classifier GANs AdaGAN — AdaGAN: Boosting Generative Models AEGAN — Learning Inverse Mapping by Autoencoder based Generative Adversarial Nets AffGAN — Amortised MAP Inference for Image Super-resolution AL-CGAN — Learning to Generate Images of Outdoor Scenes from Attributes and Semantic Layouts ALI — Adversarially Learned Inference AMGAN — Generative Adversarial Nets with Labeled Data by Activation Maximization AnoGAN — Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery ArtGAN — ArtGAN: Artwork Synthesis with Conditional Categorial GANs b-GAN — b-GAN: Unified Framework of Generative Adversarial Networks Bayesian GAN — Deep and Hierarchical Implicit Models BEGAN — BEGAN: Boundary Equilibrium Generative Adversarial Networks BiGAN — Adversarial Feature Learning BS-GAN — Boundary-Seeking Generative Adversarial Networks CGAN — Conditional Generative Adversarial Nets CCGAN — Semi-Supervised Learning with Context-Conditional Generative Adversarial Networks CatGAN — Unsupervised and Semi-supervised Learning with Categorical Generative Adversarial Networks CoGAN — Coupled Generative Adversarial Networks Context-RNN-GAN — Contextual RNN-GANs for Abstract Reasoning Diagram Generation C-RNN-GAN — C-RNN-GAN: Continuous recurrent neural networks with adversarial training CS-GAN — Improving Neural Machine Translation with Conditional Sequence Generative Adversarial Nets CVAE-GAN — CVAE-GAN: Fine-Grained Image Generation through Asymmetric Training CycleGAN — Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks DTN — Unsupervised Cross-Domain Image Generation DCGAN — Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks DiscoGAN — Learning to Discover Cross-Domain Relations with Generative Adversarial Networks DR-GAN — Disentangled Representation Learning GAN for Pose-Invariant Face Recognition DualGAN — DualGAN: Unsupervised Dual Learning for Image-to-Image Translation EBGAN — Energy-based Generative Adversarial Network f-GAN — f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization GAWWN — Learning What and Where to Draw GoGAN — Gang of GANs: Generative Adversarial Networks with Maximum Margin Ranking GP-GAN — GP-GAN: Towards Realistic High-Resolution Image Blending IAN — Neural Photo Editing with Introspective Adversarial Networks iGAN — Generative Visual Manipulation on the Natural Image Manifold IcGAN — Invertible Conditional GANs for image editing ID-CGAN- Image De-raining Using a Conditional Generative Adversarial Network Improved GAN — Improved Techniques for Training GANs InfoGAN — InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets LAGAN — Learning Particle Physics by Example: Location-Aware Generative Adversarial Networks for Physics Synthesis LAPGAN — Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks LR-GAN — LR-GAN: Layered Recursive Generative Adversarial Networks for Image Generation LSGAN — Least Squares Generative Adversarial Networks LS-GAN — Loss-Sensitive Generative Adversarial Networks on Lipschitz Densities MGAN — Precomputed Real-Time Texture Synthesis with Markovian Generative Adversarial Networks MAGAN — MAGAN: Margin Adaptation for Generative Adversarial Networks MAD-GAN — Multi-Agent Diverse Generative Adversarial Networks MalGAN — Generating Adversarial Malware Examples for Black-Box Attacks Based on GAN MaliGAN — Maximum-Likelihood Augmented Discrete Generative Adversarial Networks MARTA-GAN — Deep Unsupervised Representation Learning for Remote Sensing Images McGAN — McGan: Mean and Covariance Feature Matching GAN MDGAN — Mode Regularized Generative Adversarial Networks MedGAN — Generating Multi-label Discrete Electronic Health Records using Generative Adversarial Networks MIX+GAN — Generalization and Equilibrium in Generative Adversarial Nets (GANs) MPM-GAN — Message Passing Multi-Agent GANs MV-BiGAN — Multi-view Generative Adversarial Networks pix2pix — Image-to-Image Translation with Conditional Adversarial Networks PPGN — Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space PrGAN — 3D Shape Induction from 2D Views of Multiple Objects RenderGAN — RenderGAN: Generating Realistic Labeled Data RTT-GAN — Recurrent Topic-Transition GAN for Visual Paragraph Generation SGAN — Stacked Generative Adversarial Networks SGAN — Texture Synthesis with Spatial Generative Adversarial Networks SAD-GAN — SAD-GAN: Synthetic Autonomous Driving using Generative Adversarial Networks SalGAN — SalGAN: Visual Saliency Prediction with Generative Adversarial Networks SEGAN — SEGAN: Speech Enhancement Generative Adversarial Network SeGAN — SeGAN: Segmenting and Generating the Invisible SeqGAN — SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient SimGAN — Learning from Simulated and Unsupervised Images through Adversarial Training SketchGAN — Adversarial Training For Sketch Retrieval SL-GAN — Semi-Latent GAN: Learning to generate and modify facial images from attributes Softmax-GAN — Softmax GAN SRGAN — Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network S²GAN — Generative Image Modeling using Style and Structure Adversarial Networks SSL-GAN — Semi-Supervised Learning with Context-Conditional Generative Adversarial Networks StackGAN — StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks TGAN — Temporal Generative Adversarial Nets TAC-GAN — TAC-GAN — Text Conditioned Auxiliary Classifier Generative Adversarial Network TP-GAN — Beyond Face Rotation: Global and Local Perception GAN for Photorealistic and Identity Preserving Frontal View Synthesis Triple-GAN — Triple Generative Adversarial Nets Unrolled GAN — Unrolled Generative Adversarial Networks VGAN — Generating Videos with Scene Dynamics VGAN — Generative Adversarial Networks as Variational Training of Energy Based Models VAE-GAN — Autoencoding beyond pixels using a learned similarity metric VariGAN — Multi-View Image Generation from a Single-View ViGAN — Image Generation and Editing with Variational Info Generative AdversarialNetworks WGAN — Wasserstein GAN WGAN-GP — Improved Training of Wasserstein GANs WaterGAN — WaterGAN: Unsupervised Generative Network to Enable Real-time Color Correction of Monocular Underwater Images
32 / 33 GENERATION Explosive GrowthGAN zoo https://deephunt.in/the-gan-zoo-79597dc8c347 Explosive growth — All the named GAN variants cumulatively since 2014. Credit: Bruno Gavranović
33 / 33 GENERATION Collections of Generative ModelGAN zoo https://github.com/soumith/talks/tree/master/2017-ICCV_Venice https://github.com/hwalsuklee/tensorflow-generative-model-collections ICCV 2017 GAN Tutorial
응용 Applications
1 / 20 APPLICATIONS 힌트를 통한 이미지 변환CONDITIONAL GENERATION 생성자 Generator 스타일 변환 생성자 Generator 스케치à사진 생성자 Generator 자동 채색
2 / 20 APPLICATIONS 스타일 변환CONDITIONAL GENERATION Fast Neural Style ‘16.04 스타일
3 / 20 APPLICATIONS 스타일 변환CONDITIONAL GENERATION Artistic style transfer for videos, ‘16.10 동영상
4 / 20 APPLICATIONS 스케치 à 사진CONDITIONAL GENERATION pixel2pixel ‘16.11 (https://affinelayer.com/pixsrv/index.html) 건물 정면 사진 DB 고양이 사진 DB 구두 사진 DB 가방 사진 DB
5 / 20 APPLICATIONS 스케치 à 채색CONDITIONAL GENERATION 자동 채색 ‘17.01 (https://paintschainer.preferred.tech/)
6 / 20 APPLICATIONS 스케치 à 채색CONDITIONAL GENERATION 자동 채색 ‘17.01 (https://paintschainer.preferred.tech/)
7 / 20 APPLICATIONS 스케치 à 채색CONDITIONAL GENERATION 자동 채색 ‘17.01 (https://paintschainer.preferred.tech/) 원본 힌트 작업 시간 < 1분 정답 색상 힌트에 따라 채색
8 / 20 APPLICATIONS 실시간 사용자 입력CONDITIONAL GENERATION 색상 힌트에 따라 채색 Generative Visual Manipulation on the Natural Image Manifold , ‘16.09 ①풀 ②선택 ③산 ④선택 ⑤하늘
9 / 20 APPLICATIONS 실시간 사용자 입력CONDITIONAL GENERATION 색상 힌트에 따라 채색 Generative Visual Manipulation on the Natural Image Manifold , ‘16.09 동영상 https://youtu.be/5jfViPdYLic
10 / 20 APPLICATIONS FONT GENERATION https://kaonashi-tyc.github.io/2017/04/06/zi2zi.html , ‘17.04.06 랜덤 폰트 생성 한자 폰트 à 한글 폰트
11 / 20 APPLICATIONS FONT GENERATION http://fontto.twiiks.co/demo/ Handwritten Character to Font
12 / 20 Multimodal Feature Learner StackGAN : Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks , ‘16.12 APPLICATIONS StackGANGAN+VAE
13 / 20 Learning a Probabilistic latent Space of Object Shapes via 3D Generative-Adversarial Modeling (3D-GAN), ‘16.10 Multimodal Feature Learner APPLICATIONS 3DGANGAN+VAE
14 / 20 Denoising SEGAN: Speech Enhancement Generative Adversarial Network ‘17. 03. 28 Nothing is safe. There will be no repeat of that performance, that I can guarantee. before after before after APPLICATIONS SEGANGAN+VAE
15 / 20 Age Progression/Regression by Conditional Adversarial Autoencoder https://zzutk.github.io/Face-Aging-CAAE/ APPLICATIONS Papers in CVPR2017GAN+VAE
16 / 20 Age Progression/Regression by Conditional Adversarial Autoencoder https://zzutk.github.io/Face-Aging-CAAE/ APPLICATIONS Papers in CVPR2017GAN+VAE
17 / 20 Hallucinating Very Low-Resolution Unaligned and Noisy Face Images by Transformative Discriminative Autoencoders http://www.porikli.com/mysite/pdfs/porikli%202017%20-%20Hallucinating%20very%20low- resolution%20unaligned%20and%20noisy%20face%20images%20by%20transformative%20discriminative%20autoencoders.pdf APPLICATIONS Papers in CVPR2017GAN+VAE 16x16 à 128x128
18 / 20 Hallucinating Very Low-Resolution Unaligned and Noisy Face Images by Transformative Discriminative Autoencoders http://www.porikli.com/mysite/pdfs/porikli%202017%20-%20Hallucinating%20very%20low- resolution%20unaligned%20and%20noisy%20face%20images%20by%20transformative%20discriminative%20autoencoders.pdf APPLICATIONS Papers in CVPR2017GAN+VAE TUN loss DL loss TE loss
19 / 20 A Generative Model of People in Clothing https://arxiv.org/abs/1705.04098 APPLICATIONS Papers in ICCV2017GAN+VAE
20 / 20 The Conditional Analogy GAN: Swapping Fashion Articles on People Images http://openaccess.thecvf.com/content_ICCV_2017_workshops/w32/html/Jetchev_The_Conditional_Analogy_ICCV_2017_paper.html APPLICATIONS Papers in ICCV2017Conditional GAN
마무리 Closing
1 / 6 CLOSING 진짜 사진을 찾아보세요!FACE GENERATION DCGAN ‘15.11.19 64 x 64 pixels BEGAN ‘17.03.31 128 x 128 pixels 4배 화질 향상 https://arxiv.org/abs/1703.10717
2 / 6 CLOSING DCGAN ‘15.11.19 64 x 64 pixels Create Anime Characters with A.I. ‘17.08.14 128 x 128 pixels 4배 화질 향상 http://make.girls.moe/#/ 결과 비교CHARACTER GENERATION
3 / 6 CLOSING Deep Feature Interpolation (CVPR 2017) 1000 x 1000 pixels https://github.com/paulu/deepfeatinterp 얼굴 이미지CONDITIONAL GENERATION
4 / 6 CLOSING 진짜 사진을 찾아보세요!FACE GENERATION BEGAN ‘17.03.31 128 x 128 pixels PGGAN ‘17.10.27 https://arxiv.org/abs/1710.10196
5 / 6 CLOSING 진짜 사진을 찾아보세요!FACE GENERATION PGGAN ‘17.10.27 1024 x 1024 pixels 16*16배 화질 향상
HARDWARE 6 / 6 CLOSING 컴퓨터 비전 기술의 비약적 발전 배경 BIG DATA HARDWARE 논문 공유 코드 공유 개발 플랫폼 공유 공유의 문화
End of Document
List of Web Demos • Classification : https://www.clarifai.com/demo • Segmentation : http://www.robots.ox.ac.uk/~szheng/crfasrnndemo • Vae mnist : http://www.dpkingma.com/sgvb_mnist_demo/demo.html • Vae gray face : http://vdumoulin.github.io/morphing_faces/online_demo.html • Vae sketch-rnn : https://magenta.tensorflow.org/assets/sketch_rnn_demo/multi_vae.html • 1d gan : http://cs.stanford.edu/people/karpathy/gan/ • Dcgan asian face : https://carpedm20.github.io/faces/ • Dcgan character generation : http://mattya.github.io/chainer-DCGAN • Style-transfer : http://demos.algorithmia.com/deep-style/ • Pix2pix : https://affinelayer.com/pixsrv/index.html • Colorization : https://paintschainer.preferred.tech/ • Font generation : http://fontto.twiiks.co/demo/ • Conditional character generation : http://make.girls.moe/#/

More Related Content

PDF
오토인코더의 모든 것
byNAVER Engineering
 
PDF
자습해도 모르겠던 딥러닝, 머리속에 인스톨 시켜드립니다.
byYongho Ha
 
PDF
[RLkorea] 각잡고 로봇팔 발표
byashley ryu
 
PPTX
【LT資料】 Neural Network 素人なんだけど何とかご機嫌取りをしたい
byTakuji Tahara
 
PPTX
[DL輪読会]逆強化学習とGANs
byDeep Learning JP
 
PPTX
【DL輪読会】Dropout Reduces Underfitting
byDeep Learning JP
 
PDF
PR-409: Denoising Diffusion Probabilistic Models
byHyeongmin Lee
 
PPTX
[DL輪読会]Wavenet a generative model for raw audio
byDeep Learning JP
 
오토인코더의 모든 것
byNAVER Engineering
 
자습해도 모르겠던 딥러닝, 머리속에 인스톨 시켜드립니다.
byYongho Ha
 
[RLkorea] 각잡고 로봇팔 발표
byashley ryu
 
【LT資料】 Neural Network 素人なんだけど何とかご機嫌取りをしたい
byTakuji Tahara
 
[DL輪読会]逆強化学習とGANs
byDeep Learning JP
 
【DL輪読会】Dropout Reduces Underfitting
byDeep Learning JP
 
PR-409: Denoising Diffusion Probabilistic Models
byHyeongmin Lee
 
[DL輪読会]Wavenet a generative model for raw audio
byDeep Learning JP
 

What's hot

PPTX
딥 러닝 자연어 처리를 학습을 위한 파워포인트. (Deep Learning for Natural Language Processing)
byWON JOON YOO
 
PDF
파이썬으로 나만의 강화학습 환경 만들기
by정주 김
 
PDF
LLM 모델 기반 서비스 실전 가이드
byTae Young Lee
 
PDF
SSII2022 [SS2] 少ないデータやラベルを効率的に活用する機械学習技術 〜 足りない情報をどのように補うか？〜
bySSII
 
PDF
Wasserstein GAN 수학 이해하기 I
bySungbin Lim
 
PDF
강화 학습 기초 Reinforcement Learning an introduction
byTaehoon Kim
 
PDF
【論文読み会】Deep Clustering for Unsupervised Learning of Visual Features
byARISE analytics
 
PDF
負の二項分布について
byHiroshi Shimizu
 
PPTX
[DL輪読会]Revisiting Deep Learning Models for Tabular Data (NeurIPS 2021) 表形式デー...
byDeep Learning JP
 
PDF
RBM、Deep Learningと学習(全脳アーキテクチャ若手の会 第3回DL勉強会発表資料)
byTakuma Yagi
 
PPTX
[DL輪読会]NVAE: A Deep Hierarchical Variational Autoencoder
byDeep Learning JP
 
PPTX
ベイズ最適化によるハイパラーパラメータ探索
by西岡 賢一郎
 
PDF
[DL輪読会]Weight Agnostic Neural Networks
byDeep Learning JP
 
PPT
AlphaGo Zero 解説
bysuckgeun lee
 
PPTX
【DL輪読会】論文解説：Offline Reinforcement Learning as One Big Sequence Modeling Problem
byDeep Learning JP
 
PPTX
強化学習2章
byhiroki yamaoka
 
PDF
続分かりやすいパターン認識 4章後半(4.7以降)
byMotoya Wakiyama
 
PDF
EMアルゴリズム
bySotetsu KOYAMADA（小山田創哲）
 
PDF
[DL輪読会]近年のエネルギーベースモデルの進展
byDeep Learning JP
 
PDF
[DL輪読会]GANとエネルギーベースモデル
byDeep Learning JP
 
딥 러닝 자연어 처리를 학습을 위한 파워포인트. (Deep Learning for Natural Language Processing)
byWON JOON YOO
 
파이썬으로 나만의 강화학습 환경 만들기
by정주 김
 
LLM 모델 기반 서비스 실전 가이드
byTae Young Lee
 
SSII2022 [SS2] 少ないデータやラベルを効率的に活用する機械学習技術 〜 足りない情報をどのように補うか？〜
bySSII
 
Wasserstein GAN 수학 이해하기 I
bySungbin Lim
 
강화 학습 기초 Reinforcement Learning an introduction
byTaehoon Kim
 
【論文読み会】Deep Clustering for Unsupervised Learning of Visual Features
byARISE analytics
 
負の二項分布について
byHiroshi Shimizu
 
[DL輪読会]Revisiting Deep Learning Models for Tabular Data (NeurIPS 2021) 表形式デー...
byDeep Learning JP
 
RBM、Deep Learningと学習(全脳アーキテクチャ若手の会 第3回DL勉強会発表資料)
byTakuma Yagi
 
[DL輪読会]NVAE: A Deep Hierarchical Variational Autoencoder
byDeep Learning JP
 
ベイズ最適化によるハイパラーパラメータ探索
by西岡 賢一郎
 
[DL輪読会]Weight Agnostic Neural Networks
byDeep Learning JP
 
AlphaGo Zero 解説
bysuckgeun lee
 
【DL輪読会】論文解説：Offline Reinforcement Learning as One Big Sequence Modeling Problem
byDeep Learning JP
 
強化学習2章
byhiroki yamaoka
 
続分かりやすいパターン認識 4章後半(4.7以降)
byMotoya Wakiyama
 
EMアルゴリズム
bySotetsu KOYAMADA（小山田創哲）
 
[DL輪読会]近年のエネルギーベースモデルの進展
byDeep Learning JP
 
[DL輪読会]GANとエネルギーベースモデル
byDeep Learning JP
 

Similar to 그림 그리는 AI

PDF
[NDC2017] 딥러닝으로 게임 콘텐츠 제작하기 - VAE를 이용한 콘텐츠 생성 기법 연구 사례
byHwanhee Kim
 
PDF
GAN in medical imaging
byCheng-Bin Jin
 
PDF
Photo-realistic Single Image Super-resolution using a Generative Adversarial ...
byHansol Kang
 
PPTX
Variational Autoencoder Tutorial
byHojin Yang
 
PDF
Generative adversarial networks
byKyuri Kim
 
PDF
2021 04-01-dalle
byJAEMINJEONG5
 
PDF
AI 로봇 아티스트의 비밀(창원대학교 정보통신공학과 특강)
byChangwon National University
 
PPTX
20230213_ComputerVision_연구.pptx
byssuser7807522
 
PPTX
Generative Adversarial Networks and Their Applications in Medical Imaging
bySanghoon Hong
 
PDF
Recent advances of AI for medical imaging : Engineering perspectives
byNamkug Kim
 
PPTX
Enhancing VAEs for collaborative filtering : flexible priors & gating mechanisms
byseungwoo kim
 
PDF
Icml2018 naver review
byNAVER Engineering
 
PDF
Introduction to GAN
byJimin Lee
 
PPTX
Deep learning.pptxst8itsstitissitdyiitsistitsitd
byts24051
 
PPTX
AutoEncoder&GAN Introduction
bypko89403
 
PPTX
Reviews on Deep Generative Models in the early days / GANs & VAEs paper review
bychangedaeoh
 
PDF
[PR12] intro. to gans jaejun yoo
byJaeJun Yoo
 
PPTX
White box in Computer Vision
byJaehyuk Heo
 
PPTX
Cvpr 2018 papers review (efficient computing)
byDonghyunKang12
 
PPTX
iT Cafe - Generative Adversarial Network(GAN)
byDongmin Kim
 
[NDC2017] 딥러닝으로 게임 콘텐츠 제작하기 - VAE를 이용한 콘텐츠 생성 기법 연구 사례
byHwanhee Kim
 
GAN in medical imaging
byCheng-Bin Jin
 
Photo-realistic Single Image Super-resolution using a Generative Adversarial ...
byHansol Kang
 
Variational Autoencoder Tutorial
byHojin Yang
 
Generative adversarial networks
byKyuri Kim
 
2021 04-01-dalle
byJAEMINJEONG5
 
AI 로봇 아티스트의 비밀(창원대학교 정보통신공학과 특강)
byChangwon National University
 
20230213_ComputerVision_연구.pptx
byssuser7807522
 
Generative Adversarial Networks and Their Applications in Medical Imaging
bySanghoon Hong
 
Recent advances of AI for medical imaging : Engineering perspectives
byNamkug Kim
 
Enhancing VAEs for collaborative filtering : flexible priors & gating mechanisms
byseungwoo kim
 
Icml2018 naver review
byNAVER Engineering
 
Introduction to GAN
byJimin Lee
 
Deep learning.pptxst8itsstitissitdyiitsistitsitd
byts24051
 
AutoEncoder&GAN Introduction
bypko89403
 
Reviews on Deep Generative Models in the early days / GANs & VAEs paper review
bychangedaeoh
 
[PR12] intro. to gans jaejun yoo
byJaeJun Yoo
 
White box in Computer Vision
byJaehyuk Heo
 
Cvpr 2018 papers review (efficient computing)
byDonghyunKang12
 
iT Cafe - Generative Adversarial Network(GAN)
byDongmin Kim
 

More from NAVER Engineering

PDF
서비스 운영을 위한 디자인시스템 프로젝트
byNAVER Engineering
 
PDF
플랫폼 디자이너 없이 디자인 시스템을 구축하는 프로덕트 디자이너의 우당탕탕 고통 연대기
byNAVER Engineering
 
PDF
쏘카프레임 구축 배경과 과정
byNAVER Engineering
 
PDF
React vac pattern
byNAVER Engineering
 
PDF
진화하는 디자인 시스템(걸음마 편)
byNAVER Engineering
 
PDF
이번 생에 디자인 시스템은 처음이라
byNAVER Engineering
 
PDF
디자인 시스템에 직방 ZUIX
byNAVER Engineering
 
PDF
200820 NAVER TECH CONCERT 15_Code Review is Horse(코드리뷰는 말이야)(feat.Latte)
byNAVER Engineering
 
PDF
날고 있는 여러 비행기 넘나 들며 정비하기
byNAVER Engineering
 
PDF
BPL(Banksalad Product Language) 무야호
byNAVER Engineering
 
PDF
200820 NAVER TECH CONCERT 14_야 너두 할 수 있어. 비전공자, COBOL 개발자를 거쳐 네이버에서 FE 개발하게 된...
byNAVER Engineering
 
PDF
200819 NAVER TECH CONCERT 07_신입 iOS 개발자 개발업무 적응기
byNAVER Engineering
 
PDF
200820 NAVER TECH CONCERT 11_빠르게 성장하는 슈퍼루키로 거듭나기
byNAVER Engineering
 
PDF
200820 NAVER TECH CONCERT 12_상반기 네이버 인턴을 돌아보며
byNAVER Engineering
 
PDF
200820 NAVER TECH CONCERT 13_네이버에서 오픈 소스 개발을 통해 성장하는 방법
byNAVER Engineering
 
PDF
200819 NAVER TECH CONCERT 05_모르면 손해보는 Android 디버깅/분석 꿀팁 대방출
byNAVER Engineering
 
PDF
200819 NAVER TECH CONCERT 08_성능을 고민하는 슬기로운 개발자 생활
byNAVER Engineering
 
PDF
200819 NAVER TECH CONCERT 09_Case.xcodeproj - 좋은 동료로 거듭나기 위한 노하우
byNAVER Engineering
 
PDF
200819 NAVER TECH CONCERT 03_화려한 코루틴이 내 앱을 감싸네! 코루틴으로 작성해보는 깔끔한 비동기 코드
byNAVER Engineering
 
PDF
200819 NAVER TECH CONCERT 10_맥북에서도 아이맥프로에서 빌드하는 것처럼 빌드 속도 빠르게 하기
byNAVER Engineering
 
서비스 운영을 위한 디자인시스템 프로젝트
byNAVER Engineering
 
플랫폼 디자이너 없이 디자인 시스템을 구축하는 프로덕트 디자이너의 우당탕탕 고통 연대기
byNAVER Engineering
 
쏘카프레임 구축 배경과 과정
byNAVER Engineering
 
React vac pattern
byNAVER Engineering
 
진화하는 디자인 시스템(걸음마 편)
byNAVER Engineering
 
이번 생에 디자인 시스템은 처음이라
byNAVER Engineering
 
디자인 시스템에 직방 ZUIX
byNAVER Engineering
 
200820 NAVER TECH CONCERT 15_Code Review is Horse(코드리뷰는 말이야)(feat.Latte)
byNAVER Engineering
 
날고 있는 여러 비행기 넘나 들며 정비하기
byNAVER Engineering
 
BPL(Banksalad Product Language) 무야호
byNAVER Engineering
 
200820 NAVER TECH CONCERT 14_야 너두 할 수 있어. 비전공자, COBOL 개발자를 거쳐 네이버에서 FE 개발하게 된...
byNAVER Engineering
 
200819 NAVER TECH CONCERT 07_신입 iOS 개발자 개발업무 적응기
byNAVER Engineering
 
200820 NAVER TECH CONCERT 11_빠르게 성장하는 슈퍼루키로 거듭나기
byNAVER Engineering
 
200820 NAVER TECH CONCERT 12_상반기 네이버 인턴을 돌아보며
byNAVER Engineering
 
200820 NAVER TECH CONCERT 13_네이버에서 오픈 소스 개발을 통해 성장하는 방법
byNAVER Engineering
 
200819 NAVER TECH CONCERT 05_모르면 손해보는 Android 디버깅/분석 꿀팁 대방출
byNAVER Engineering
 
200819 NAVER TECH CONCERT 08_성능을 고민하는 슬기로운 개발자 생활
byNAVER Engineering
 
200819 NAVER TECH CONCERT 09_Case.xcodeproj - 좋은 동료로 거듭나기 위한 노하우
byNAVER Engineering
 
200819 NAVER TECH CONCERT 03_화려한 코루틴이 내 앱을 감싸네! 코루틴으로 작성해보는 깔끔한 비동기 코드
byNAVER Engineering
 
200819 NAVER TECH CONCERT 10_맥북에서도 아이맥프로에서 빌드하는 것처럼 빌드 속도 빠르게 하기
byNAVER Engineering
 

Recently uploaded

PDF
threat-hunters-cookbook for cybersecurity
bylobster6719
 
PDF
IAC 500 Sensor - Humidity Measurement Device
byCS Instruments
 
PDF
Xemelgo - RFID Industry Predictions for 2026
byRich Rogers
 
PPT
HDTV and DTV Standards: The United States Opts for a Digital HDTV Standard
byJeffrey Hart
 
PPTX
Scrape YouTube Video Data for Influencer Analytics.pptx
byWeb Data Crawler
 
PDF
What is Voice User Interface (VUI) Definition & Examples.pdf
byDesign Studio UI UX
 
PDF
Video Infrastructure_ Streaming Architecture and Delivery Systems.pdf
byTenbyte Limited
 
PPTX
Digital transformation success powered by EPM and NexInfo.pptx
byjayasuryanexinfo
 
PDF
Chapter 1 Introduction to Computer Security.pdf
byGetnet Tigabie Askale -(GM)
 
PDF
Regenerative Agriculture Finance : Environmental impact
byrose mason
 
PDF
The map to conquer linear algebra for IT engineer
byakipii ogaoga
 
PPTX
AN Introduction to UNIX File System—An Approach
bysonjuktaweb
 
PPTX
Information about Trusted Platform Module(TPM)
bynewtoknow techs
 
PPTX
Cesium formate brine - A brief history - prepared by John Downs in May 2011
byJohn Downs
 
PDF
Azure Identity, Governance, and Monitoring solutions
byVICTOR MAESTRE RAMIREZ
 
PPTX
AI Bot Traffic Surge: Retail Fraud Threat for Age-Restricted Websites
byFTx Identity
 
PDF
Artificial Intelligence and Barbarism - Conceptual Map
byalfadix
 
PPTX
UNIT III TYPOLOGY OF CRIME AND CRIMINAL BEHAVIOUR (1).pptx
bybloodyhumanoid
 
PPT
Information and Communication Technologies and Power
byJeffrey Hart
 
PDF
Digital Marketing Trends in 2026: AI, Data, Content & Future Growth
byConacent Consulting
 
threat-hunters-cookbook for cybersecurity
bylobster6719
 
IAC 500 Sensor - Humidity Measurement Device
byCS Instruments
 
Xemelgo - RFID Industry Predictions for 2026
byRich Rogers
 
HDTV and DTV Standards: The United States Opts for a Digital HDTV Standard
byJeffrey Hart
 
Scrape YouTube Video Data for Influencer Analytics.pptx
byWeb Data Crawler
 
What is Voice User Interface (VUI) Definition & Examples.pdf
byDesign Studio UI UX
 
Video Infrastructure_ Streaming Architecture and Delivery Systems.pdf
byTenbyte Limited
 
Digital transformation success powered by EPM and NexInfo.pptx
byjayasuryanexinfo
 
Chapter 1 Introduction to Computer Security.pdf
byGetnet Tigabie Askale -(GM)
 
Regenerative Agriculture Finance : Environmental impact
byrose mason
 
The map to conquer linear algebra for IT engineer
byakipii ogaoga
 
AN Introduction to UNIX File System—An Approach
bysonjuktaweb
 
Information about Trusted Platform Module(TPM)
bynewtoknow techs
 
Cesium formate brine - A brief history - prepared by John Downs in May 2011
byJohn Downs
 
Azure Identity, Governance, and Monitoring solutions
byVICTOR MAESTRE RAMIREZ
 
AI Bot Traffic Surge: Retail Fraud Threat for Age-Restricted Websites
byFTx Identity
 
Artificial Intelligence and Barbarism - Conceptual Map
byalfadix
 
UNIT III TYPOLOGY OF CRIME AND CRIMINAL BEHAVIOUR (1).pptx
bybloodyhumanoid
 
Information and Communication Technologies and Power
byJeffrey Hart
 
Digital Marketing Trends in 2026: AI, Data, Content & Future Growth
byConacent Consulting
 

그림 그리는 AI

  • 1.
  • 2.
    인식 이미지 내의 존재하는정보를 찾는 기술 얼굴 인식 홍채 인식 번호판 인식 지문 인식 생성 특정 정보를 담는 이미지를 생성하는 기술 스타일 변환스타일 이미지 컴퓨터 비전 기술INTRODUCTION
  • 3.
  • 4.
    ü ImageNet 대회 •1000 종류의 클래스 존재 u 목표 : 이미지 내 존재하는 객체 분류 < 사람=5% u 2015년부터 사람보다 잘 하기 시작 (2017년으로 종료) 빅데이터+딥러닝 > 사람 1 / 6 RECOGNITION 대회 현황IMAGENET 사람이 5% 오차? : 정답의 오류 및 종류의 세분화
  • 5.
    정해진 수의 물건들의특징을 잘 파악하여 구분할 수 있다! 딥뉴럴 네트워크 개와 고양이를 구분하는 모델 RECOGNITION 2015년부터 AI가 잘 할 수 있게 된 것은?IMAGENET 2 / 6
  • 6.
    RECOGNITION 이미지 태그 생성APPLICATION 이미지와관련 있는 태그 찾기 https://www.clarifai.com/demo 3 / 6
  • 7.
    RECOGNITION 객체 분류APPLICATION ü 사각형영역 단위로 구분 (RCNN, ‘15.04) ü 화소 단위로 구분 (DeepLab, ‘17.03) 4 / 6
  • 8.
    RECOGNITION 신체 부위 분류APPLICATION üOpenpose, ‘17.04 오승환 투구 슬로우 모션 영상 https://github.com/CMU-Perceptual-Computing-Lab/openpose 5 / 6 얼굴 + 신체 + 손
  • 9.
  • 10.
    딥뉴럴 네트워크 개와 고양이를생성할 수 있는 모델 분류 모델 보다는 개와 고양이를 제대로 이해하고 있다 Generative Model “What I cannot create, I do not understand.” - Richard Feynman 1 / 13 GENERATION 2015년부터 AI 연구자들이 관심 갖게 된 것은?INTORUDCTION
  • 11.
    StackGAN : Textto Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks , ‘16.12 AI가 생성한 사진을 찾아보세요! (1) This flower has overlapping pink pointed petals surrounding a ring of short yellow filaments 이 꽃은 짧은 노란색 필라멘트의 고리를 둘러싼 핑크색 뾰족한 꽃잎이 겹쳐져 있습니다 (2) This flower has upturned petals which are thin and orange with rounded edges 이 꽃은 둥근 모서리를 가진 얇고 오렌지색의 꽃잎이 위로 향해 있습니다 (3) A flower with small pink petals and a massive central orange and black stamen cluster 작은 분홍색 꽃잎들과 중심에 다수의 오렌지색과 검은 색 수술 군집이 있는 꽃 (1) (2) (3) 2 / 13 GENERATION Quiz : stackGANINTORUDCTION
  • 12.
  • 13.
  • 14.
    5 / 13 GENERATION Generative ModelsINTORUDCTION Taxomomy http://www.iangoodfellow.com/slides/2016-12-04-NIPS.pdf 분포에대한 모델을 가정하고 그 모델을 결정하는 파라미터를 조절해 가며 타겟 분포가 되도록 노력 생성되는 샘플들의 분포가 타겟 분포가 되도록 노력 VAE GAN
  • 15.
  • 16.
    Variational • 𝒑(𝒙)를구할 때 variational learning/inference 활용 AutoEncoder 1 / 33 GENERATION 용어의 의미VAE Auto-Encoding Variational Bayes , ‘14.03 Latent Variable Target Data 𝑥 𝑔)* 𝑧 𝑔)+ 𝑧 𝑝 𝑥 𝑔)+ 𝑧 < 𝑝 𝑥 𝑔)* 𝑧 확률 분포를 정해주는 파라미터를 추정. (예. 가우시안일 경우 평균,표준편차) 𝑧 𝑔)(𝑧)Generator gθ(.) 𝑝 𝑧 𝑝 𝑥 𝑔)(𝑧) sampling modeling - 𝑝 𝑥 𝑔)(𝑧) 𝑝 𝑧 𝑑𝑧 = 𝑝(𝑥) variational learning - 𝑝 𝑥 𝑔)(𝑧) 𝑝 𝑧 𝑑𝑧 = 𝐸2(3)[𝑝)(𝑥|𝑧)]
  • 17.
    Variational • 𝒑(𝒙)를구할 때 variational learning/inference 활용 AutoEncoder 2 / 33 GENERATION 용어의 의미VAE Auto-Encoding Variational Bayes , ‘14.03 𝑧 𝑔)(𝑧)Generator gθ(.) 𝑝 𝑧 sampling • 입력과 출력이 동일한 네트워크 𝑝 𝑧|𝑥 x를 잘 생성해는 z가 샘플링 잘 되도록, x를 힌트로 제공함 𝑝 𝑧|𝑞8(𝑥) p(z|x)의 모델을 가정하고, 그 모델을 결정짓는 파라미터를 네트워크로 추정 Decoder gθ(.) Encoder qφ(.) 𝑥 𝑝 𝑧|𝑞8(𝑥) 𝑞8(𝑥) sampling 𝑧 𝑔) 𝑧 𝑥→
  • 18.
    3 / 33 GENERATION VariationalInferenceVAE Variational Lower Bound (Evidence Lower BOund, ELBO) log 𝑝 𝑥 = 𝐸𝐿𝐵𝑂 + 𝐾𝐿(𝑞8 𝑧|𝑥 ∥ 𝑝 𝑧|𝑥 ) log 𝑝 𝑥 𝐾𝐿(𝑞8 𝑧|𝑥 ∥ 𝑝 𝑧|𝑥 ) 𝐸𝐿𝐵𝑂(𝜙) 𝜙D 𝜙E Optimization Problem 1 on 𝜙: Variational Inference argmin 8 𝐾𝐿(𝑞8 𝑧|𝑥 ∥ 𝑝 𝑧|𝑥 ) == argmax 8 𝐸𝐿𝐵𝑂(𝜙) log 𝑝 𝑥 ≥ 𝔼NO 3|P log 𝑝) 𝑥|𝑧 − 𝐾𝐿(𝑞8 𝑧|𝑥 ∥ 𝑝 𝑧 ) = 𝐸𝐿𝐵𝑂 Optimization Problem 2 on 𝜃: Variational Learning argmax ) 𝔼NO 3|P log 𝑝) 𝑥|𝑧 = argmax ) 𝐸𝐿𝐵𝑂(𝜃) Final Optimization Problem argmax ),8 𝐸𝐿𝐵𝑂(𝜃, 𝜙) • KL : Kullback-Leibler Divergence 두 확률 분포가 얼마나 다른지에 대한 척도
  • 19.
    𝐿U 𝜙, 𝜃,𝑥U = −𝔼NO 3|P log 𝑝) 𝑥U|𝑧 + 𝐾𝐿(𝑞8 𝑧|𝑥U ∥ 𝑝 𝑧 ) Reconstruction Error 복원 오차 입력과 출력 간의 cross-entropy Regularization 제약 조건 Prior분포와의 다른 정도 [ VAE의 특징들 ] 1. Decoder가 최소한 학습 데이터는 생성해 낼 수 있게 된다. à 생성된 데이터가 학습 데이터 좀 닮아 있다. 2. Encoder가 최소한 학습 데이터는 잘 latent vector로 표현할 수 있게 된다. à 데이터의 추상화를 위해 많이 사용된다. 4 / 33 GENERATION Encoder & DecoderVAE 𝑥U Encoder 𝑞8(𝑧|𝑥U) 𝜇U 𝜎U 𝜖~𝑁(0, 𝐼) 𝜖U 𝑧U Decoder 𝑝) 𝑥|𝑧 𝑦U = 𝑥U Bernoulli Gaussian Cost Function
  • 20.
    28 𝑥U 𝑝U 𝐷 28 𝐷=784 MLP with 2 hidden layers (500, 500) Architecture 5/ 33 GENERATION Result : MNISTVAE 𝑔)(_)𝑞8 _ 𝑞8 𝑧|𝑥 ~𝑧 𝑔) 𝑥|𝑧 Encoder Posterior Inference Network Decoder Generator Generation Network SAMPLING 𝜇U 𝜎U Latent Space
  • 21.
    Reproduce Input image |z| =2 |z| =5|z| =20 https://github.com/hwalsuklee/tensorflow-mnist-VAE 6 / 33 GENERATION Result : MNISTVAE
  • 22.
  • 23.
    Learned Manifold 학습이 잘 되었을수록 2D공간에서 같은 숫자들을 생성하는 z들은 뭉쳐있고, 다른 숫자들은 생성하는 z들은 떨어져 있어야 한다. https://github.com/hwalsuklee/tensorflow-mnist-VAE 8 / 33 GENERATION Result : MNISTVAE Decoding / GenerationEncoding / Feature Extraction
  • 24.
    9 / 33 GENERATION DEMO : MNIST / Gray FaceVAE http://www.dpkingma.com/sgvb_mnist_demo/demo.html |z|=12 24 24 HandwrittenDigits Generation http://vdumoulin.github.io/morphing_faces/online_demo.html |z|=29 64 64 Gray Face Generation
  • 25.
  • 26.
  • 27.
  • 28.
  • 29.
    Handwriting styles obtained by fixing the class label and varying z |z| = 2 y=[1,0,0,0,0,0,0,0,0,0] y=[0,1,0,0,0,0,0,0,0,0]y=[0,0,1,0,0,0,0,0,0,0] y=[0,0,0,1,0,0,0,0,0,0] y=[0,0,0,0,1,0,0,0,0,0] y=[0,0,0,0,0,0,1,0,0,0]y=[0,0,0,0,0,1,0,0,0,0] y=[0,0,0,0,0,0,0,0,1,0]y=[0,0,0,0,0,0,0,1,0,0] y=[0,0,0,0,0,0,0,0,0,1] MNIST resultsCVAE https://github.com/hwalsuklee/tensorflow-mnist-CVAE 14 / 33 GENERATION
  • 30.
    Z-sampling 각 행 별로,고정된 z값에 대해서 label정보만 바꿔 서 이미지 생성 (스타일 유지하면 숫자만 바뀜) MNIST resultsCVAE Analogies : Result in paper Semi-Supervised Learning with Deep Generative Models : https://arxiv.org/abs/1406.5298 15 / 33 GENERATION
  • 31.
    𝑧D 𝑧E 𝑧` 𝑧a 𝑐c 𝑐D 𝑐E𝑐` 𝑐a 𝑐d 𝑐e 𝑐f 𝑐g 𝑐h Handwriting style for a given z must be preserved for all labels Analogies |z| = 2 MNIST resultsCVAE https://github.com/hwalsuklee/tensorflow-mnist-CVAE 𝑐c 𝑐D 𝑐E 𝑐` 𝑐a 𝑐d 𝑐e 𝑐f 𝑐g 𝑐h Real handwritten image 실제로 손으로 쓴 글씨 ‘3’을 CVAE의 label정보와 같이 넣었을 때 얻는 latent vector는 decoder의 고정 입력으로 하고, label정보만 바꿨을 경우 16 / 33 GENERATION
  • 32.
  • 33.
    17 / 33 GENERATION 용어의의미GAN Generative Adversarial Networks , ‘14.06 Generative • Generator (생성자) 및 Discriminator가 (구분자) 서로 적대적으로 행동 • 생성 모델 Adversarial Networks • 생성자 및 구분자 모두 딥뉴럴 네트워크로 모델링 x 0~1D G G(z) Discriminator Generator z “The coolest idea in ML in the last twenty years” – Yann LeCun
  • 34.
    생성자 Generator 구별자 Discriminator G A Networks 두 개의네트워크 dversarial 적대적으로 학습된 enerative 데이터 생성목적으로 실제 / 가짜 가짜 사진 실제 데이터 실제 사진 구별자는 생성자의 결과를 거짓이라고 판단하려고 노력 생성자는 구별자가 참이라고 판단하게 하려고 노력 거짓말 쟁이 거짓말 탐지기 적대적 18 / 33 GENERATION 꽃 사진 생성 예시GAN
  • 35.
    Noise (Latent variable) 생성자 Generator 구별자 Discriminator 실 데이터생성자 Yes / No 실 데이터? Generator가 매 번 다른 샘플을 만들어 내기 위한 입력 정답 얼굴 DB 구별자 성능 향상 생성자 성능 향상 19 / 33 GENERATION 얼굴 사진 생성 예시GAN
  • 36.
    Noise (Latent variable) eneratorDiscriminator Data sample Yes / No G 𝑝ijkj(𝑥) 𝐷 𝑥 =1 𝐷 𝐺 𝑧 = 0 𝑥 𝐺(𝑧) 𝑧 𝐷 𝐺 𝑧 =1 𝑝3(𝑧) 𝑉 𝐷, 𝐺 = 𝔼P~2opqp(P) log𝐷(𝑥) + 𝔼3~2r(3) log 1 − 𝐷 𝐺 𝑧Value function of GAN Goal 𝐷∗ , 𝐺∗ = min t max u 𝑉 𝐷, 𝐺 minmax problem!!! 20 / 33 GENERATION ProblemGAN Notation
  • 37.
    D max u 𝑉 𝐷,𝐺 = max u 𝔼P~2opqp log𝐷(𝑥) + 𝔼3~2r log 1 − 𝐷 𝐺 𝑧 G is fixed Maximize prob. of D(real) Minimize prob. of D(fake) G min t 𝑉 𝐷, 𝐺 = min t 𝔼3~2r log 1 − 𝐷 𝐺 𝑧 D is fixed Maximize prob. of D(fake) = max t 𝔼3~2r log 𝐷 𝐺 𝑧 Alternating Optimization 21 / 33 GENERATION ProblemGAN Optimization 𝑉 𝐷, 𝐺 = 𝔼P~2opqp(P) log𝐷(𝑥) + 𝔼3~2r(3) log 1 − 𝐷 𝐺 𝑧Value function of GAN Goal 𝐷∗ , 𝐺∗ = min t max u 𝑉 𝐷, 𝐺
  • 38.
    𝐷∗ 𝑥 = 𝑝ijkj(𝑥) 𝑝ijkj 𝑥+ 𝑝v(𝑥)D 𝑝v ∗ 𝑥 = 𝑝ijkj(𝑥) G D(x) p_data p_g(x) Optimization for D(x) Optimization for G(z) Equilibrium : D(x)=0.5 G(z) maps z to x 1D Gaussian Approximation example 22 / 33 GENERATION ProblemGAN Optimal Solution http://cs.stanford.edu/people/karpathy/gan/ Uniform Distribution https://github.com/hwalsuklee/tensorflow-GAN-1d-gaussian-ex
  • 39.
    23 / 33 GENERATION VAE VS GANGAN ModelOptimization Image Quality Generalization VAE • Stochastic gradient descent • Converge to local minimum • Easier • Smooth • Blurry • Tend to remember input images GAN • Alternating stochastic gradient descent • Converge to saddle points • Harder v Model collapsing v Unstable convergence • Sharp • Artifact • Generate new unseen images http://aliensunmin.github.io/project/accv16tutorial/media/generative.pdf z DEx x DecoderEncoder x 0~1D G G(z) Discriminator Generator z VAE GAN
  • 40.
    Comparison between VAE vs GAN VAE : maximumlikelihood approach GAN http://videolectures.net/site/normal_dl/tag=1129740/deeplearning2017_courville_generative_models_01.pdf 24 / 33 GENERATION VAE VS GANGAN
  • 41.
    이후로 소개되는 모든논문들은 DCGAN에서 크게 벗어나지 않음 대부분의 상황에서 안정적으로 학습이 되는 Deep Convolutional GAN구조를 제안 Generator Discriminator Pooling Layers Not used. But use strided convolutions instead. Batch normalization Use except output layer Use except input layer Fully connected hidden layers Not used Activation function ReLU for all layers except for the output, which uses Tanh LeakyReLU for all layers 25 / 33 GENERATION Key ContributionDCGAN Unsupervised Representation learning with Deep Convolutional Generative Adversarial Networks (DCGAN), ‘15.11
  • 42.
    26 / 33 GENERATION 네트워크구조DCGAN 생성자 Generator 구별자 Discriminator
  • 43.
  • 44.
    28 / 33 GENERATION Face Generation Results : InterpolationDCGAN z1z2(z1, z2)를 보간 후 새로 생성한 z에 대한 결과
  • 45.
    z1 : 썬글라스를낀 남자들을 발생시키는 z값 평균 z2 : 썬글라스를 안 낀 남자들을 발생시키는 z값 평균 z3 : 썬글라스를 안 낀 여자들을 발생시키는 z값 평균 z1 z2 z3 z1-z2+z3 썬글라스를 낀 여자들 29 / 33 GENERATION Face Generation Results : Vector arithmeticDCGAN
  • 46.
  • 47.
    31 / 33 GENERATION A List of Some PapersGANzoo https://deephunt.in/the-gan-zoo-79597dc8c347 GAN — Generative Adversarial Networks 3D-GAN — Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling acGAN — Face Aging With Conditional Generative Adversarial Networks AC-GAN — Conditional Image Synthesis With Auxiliary Classifier GANs AdaGAN — AdaGAN: Boosting Generative Models AEGAN — Learning Inverse Mapping by Autoencoder based Generative Adversarial Nets AffGAN — Amortised MAP Inference for Image Super-resolution AL-CGAN — Learning to Generate Images of Outdoor Scenes from Attributes and Semantic Layouts ALI — Adversarially Learned Inference AMGAN — Generative Adversarial Nets with Labeled Data by Activation Maximization AnoGAN — Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery ArtGAN — ArtGAN: Artwork Synthesis with Conditional Categorial GANs b-GAN — b-GAN: Unified Framework of Generative Adversarial Networks Bayesian GAN — Deep and Hierarchical Implicit Models BEGAN — BEGAN: Boundary Equilibrium Generative Adversarial Networks BiGAN — Adversarial Feature Learning BS-GAN — Boundary-Seeking Generative Adversarial Networks CGAN — Conditional Generative Adversarial Nets CCGAN — Semi-Supervised Learning with Context-Conditional Generative Adversarial Networks CatGAN — Unsupervised and Semi-supervised Learning with Categorical Generative Adversarial Networks CoGAN — Coupled Generative Adversarial Networks Context-RNN-GAN — Contextual RNN-GANs for Abstract Reasoning Diagram Generation C-RNN-GAN — C-RNN-GAN: Continuous recurrent neural networks with adversarial training CS-GAN — Improving Neural Machine Translation with Conditional Sequence Generative Adversarial Nets CVAE-GAN — CVAE-GAN: Fine-Grained Image Generation through Asymmetric Training CycleGAN — Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks DTN — Unsupervised Cross-Domain Image Generation DCGAN — Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks DiscoGAN — Learning to Discover Cross-Domain Relations with Generative Adversarial Networks DR-GAN — Disentangled Representation Learning GAN for Pose-Invariant Face Recognition DualGAN — DualGAN: Unsupervised Dual Learning for Image-to-Image Translation EBGAN — Energy-based Generative Adversarial Network f-GAN — f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization GAWWN — Learning What and Where to Draw GoGAN — Gang of GANs: Generative Adversarial Networks with Maximum Margin Ranking GP-GAN — GP-GAN: Towards Realistic High-Resolution Image Blending IAN — Neural Photo Editing with Introspective Adversarial Networks iGAN — Generative Visual Manipulation on the Natural Image Manifold IcGAN — Invertible Conditional GANs for image editing ID-CGAN- Image De-raining Using a Conditional Generative Adversarial Network Improved GAN — Improved Techniques for Training GANs InfoGAN — InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets LAGAN — Learning Particle Physics by Example: Location-Aware Generative Adversarial Networks for Physics Synthesis LAPGAN — Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks LR-GAN — LR-GAN: Layered Recursive Generative Adversarial Networks for Image Generation LSGAN — Least Squares Generative Adversarial Networks LS-GAN — Loss-Sensitive Generative Adversarial Networks on Lipschitz Densities MGAN — Precomputed Real-Time Texture Synthesis with Markovian Generative Adversarial Networks MAGAN — MAGAN: Margin Adaptation for Generative Adversarial Networks MAD-GAN — Multi-Agent Diverse Generative Adversarial Networks MalGAN — Generating Adversarial Malware Examples for Black-Box Attacks Based on GAN MaliGAN — Maximum-Likelihood Augmented Discrete Generative Adversarial Networks MARTA-GAN — Deep Unsupervised Representation Learning for Remote Sensing Images McGAN — McGan: Mean and Covariance Feature Matching GAN MDGAN — Mode Regularized Generative Adversarial Networks MedGAN — Generating Multi-label Discrete Electronic Health Records using Generative Adversarial Networks MIX+GAN — Generalization and Equilibrium in Generative Adversarial Nets (GANs) MPM-GAN — Message Passing Multi-Agent GANs MV-BiGAN — Multi-view Generative Adversarial Networks pix2pix — Image-to-Image Translation with Conditional Adversarial Networks PPGN — Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space PrGAN — 3D Shape Induction from 2D Views of Multiple Objects RenderGAN — RenderGAN: Generating Realistic Labeled Data RTT-GAN — Recurrent Topic-Transition GAN for Visual Paragraph Generation SGAN — Stacked Generative Adversarial Networks SGAN — Texture Synthesis with Spatial Generative Adversarial Networks SAD-GAN — SAD-GAN: Synthetic Autonomous Driving using Generative Adversarial Networks SalGAN — SalGAN: Visual Saliency Prediction with Generative Adversarial Networks SEGAN — SEGAN: Speech Enhancement Generative Adversarial Network SeGAN — SeGAN: Segmenting and Generating the Invisible SeqGAN — SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient SimGAN — Learning from Simulated and Unsupervised Images through Adversarial Training SketchGAN — Adversarial Training For Sketch Retrieval SL-GAN — Semi-Latent GAN: Learning to generate and modify facial images from attributes Softmax-GAN — Softmax GAN SRGAN — Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network S²GAN — Generative Image Modeling using Style and Structure Adversarial Networks SSL-GAN — Semi-Supervised Learning with Context-Conditional Generative Adversarial Networks StackGAN — StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks TGAN — Temporal Generative Adversarial Nets TAC-GAN — TAC-GAN — Text Conditioned Auxiliary Classifier Generative Adversarial Network TP-GAN — Beyond Face Rotation: Global and Local Perception GAN for Photorealistic and Identity Preserving Frontal View Synthesis Triple-GAN — Triple Generative Adversarial Nets Unrolled GAN — Unrolled Generative Adversarial Networks VGAN — Generating Videos with Scene Dynamics VGAN — Generative Adversarial Networks as Variational Training of Energy Based Models VAE-GAN — Autoencoding beyond pixels using a learned similarity metric VariGAN — Multi-View Image Generation from a Single-View ViGAN — Image Generation and Editing with Variational Info Generative AdversarialNetworks WGAN — Wasserstein GAN WGAN-GP — Improved Training of Wasserstein GANs WaterGAN — WaterGAN: Unsupervised Generative Network to Enable Real-time Color Correction of Monocular Underwater Images
  • 48.
    32 / 33 GENERATION Explosive GrowthGANzoo https://deephunt.in/the-gan-zoo-79597dc8c347 Explosive growth — All the named GAN variants cumulatively since 2014. Credit: Bruno Gavranović
  • 49.
    33 / 33 GENERATION Collections of Generative ModelGANzoo https://github.com/soumith/talks/tree/master/2017-ICCV_Venice https://github.com/hwalsuklee/tensorflow-generative-model-collections ICCV 2017 GAN Tutorial
  • 50.
  • 51.
    1 / 20 APPLICATIONS 힌트를통한 이미지 변환CONDITIONAL GENERATION 생성자 Generator 스타일 변환 생성자 Generator 스케치à사진 생성자 Generator 자동 채색
  • 52.
    2 / 20 APPLICATIONS 스타일변환CONDITIONAL GENERATION Fast Neural Style ‘16.04 스타일
  • 53.
    3 / 20 APPLICATIONS 스타일변환CONDITIONAL GENERATION Artistic style transfer for videos, ‘16.10 동영상
  • 54.
    4 / 20 APPLICATIONS 스케치à 사진CONDITIONAL GENERATION pixel2pixel ‘16.11 (https://affinelayer.com/pixsrv/index.html) 건물 정면 사진 DB 고양이 사진 DB 구두 사진 DB 가방 사진 DB
  • 55.
    5 / 20 APPLICATIONS 스케치à 채색CONDITIONAL GENERATION 자동 채색 ‘17.01 (https://paintschainer.preferred.tech/)
  • 56.
    6 / 20 APPLICATIONS 스케치à 채색CONDITIONAL GENERATION 자동 채색 ‘17.01 (https://paintschainer.preferred.tech/)
  • 57.
    7 / 20 APPLICATIONS 스케치à 채색CONDITIONAL GENERATION 자동 채색 ‘17.01 (https://paintschainer.preferred.tech/) 원본 힌트 작업 시간 < 1분 정답 색상 힌트에 따라 채색
  • 58.
    8 / 20 APPLICATIONS 실시간사용자 입력CONDITIONAL GENERATION 색상 힌트에 따라 채색 Generative Visual Manipulation on the Natural Image Manifold , ‘16.09 ①풀 ②선택 ③산 ④선택 ⑤하늘
  • 59.
    9 / 20 APPLICATIONS 실시간사용자 입력CONDITIONAL GENERATION 색상 힌트에 따라 채색 Generative Visual Manipulation on the Natural Image Manifold , ‘16.09 동영상 https://youtu.be/5jfViPdYLic
  • 60.
    10 / 20 APPLICATIONS FONTGENERATION https://kaonashi-tyc.github.io/2017/04/06/zi2zi.html , ‘17.04.06 랜덤 폰트 생성 한자 폰트 à 한글 폰트
  • 61.
    11 / 20 APPLICATIONS FONTGENERATION http://fontto.twiiks.co/demo/ Handwritten Character to Font
  • 62.
    12 / 20 Multimodal Feature Learner StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks , ‘16.12 APPLICATIONS StackGANGAN+VAE
  • 63.
    13 / 20 Learninga Probabilistic latent Space of Object Shapes via 3D Generative-Adversarial Modeling (3D-GAN), ‘16.10 Multimodal Feature Learner APPLICATIONS 3DGANGAN+VAE
  • 64.
    14 / 20 Denoising SEGAN:Speech Enhancement Generative Adversarial Network ‘17. 03. 28 Nothing is safe. There will be no repeat of that performance, that I can guarantee. before after before after APPLICATIONS SEGANGAN+VAE
  • 65.
  • 66.
  • 67.
  • 68.
  • 69.
  • 70.
  • 71.
  • 72.
    1 / 6 CLOSING 진짜사진을 찾아보세요!FACE GENERATION DCGAN ‘15.11.19 64 x 64 pixels BEGAN ‘17.03.31 128 x 128 pixels 4배 화질 향상 https://arxiv.org/abs/1703.10717
  • 73.
    2 / 6 CLOSING DCGAN‘15.11.19 64 x 64 pixels Create Anime Characters with A.I. ‘17.08.14 128 x 128 pixels 4배 화질 향상 http://make.girls.moe/#/ 결과 비교CHARACTER GENERATION
  • 74.
    3 / 6 CLOSING DeepFeature Interpolation (CVPR 2017) 1000 x 1000 pixels https://github.com/paulu/deepfeatinterp 얼굴 이미지CONDITIONAL GENERATION
  • 75.
    4 / 6 CLOSING 진짜사진을 찾아보세요!FACE GENERATION BEGAN ‘17.03.31 128 x 128 pixels PGGAN ‘17.10.27 https://arxiv.org/abs/1710.10196
  • 76.
    5 / 6 CLOSING 진짜사진을 찾아보세요!FACE GENERATION PGGAN ‘17.10.27 1024 x 1024 pixels 16*16배 화질 향상
  • 77.
    HARDWARE 6 / 6 CLOSING 컴퓨터비전 기술의 비약적 발전 배경 BIG DATA HARDWARE 논문 공유 코드 공유 개발 플랫폼 공유 공유의 문화
  • 78.
  • 79.
    List of WebDemos • Classification : https://www.clarifai.com/demo • Segmentation : http://www.robots.ox.ac.uk/~szheng/crfasrnndemo • Vae mnist : http://www.dpkingma.com/sgvb_mnist_demo/demo.html • Vae gray face : http://vdumoulin.github.io/morphing_faces/online_demo.html • Vae sketch-rnn : https://magenta.tensorflow.org/assets/sketch_rnn_demo/multi_vae.html • 1d gan : http://cs.stanford.edu/people/karpathy/gan/ • Dcgan asian face : https://carpedm20.github.io/faces/ • Dcgan character generation : http://mattya.github.io/chainer-DCGAN • Style-transfer : http://demos.algorithmia.com/deep-style/ • Pix2pix : https://affinelayer.com/pixsrv/index.html • Colorization : https://paintschainer.preferred.tech/ • Font generation : http://fontto.twiiks.co/demo/ • Conditional character generation : http://make.girls.moe/#/