일 | 월 | 화 | 수 | 목 | 금 | 토 |
---|---|---|---|---|---|---|
1 | ||||||
2 | 3 | 4 | 5 | 6 | 7 | 8 |
9 | 10 | 11 | 12 | 13 | 14 | 15 |
16 | 17 | 18 | 19 | 20 | 21 | 22 |
23 | 24 | 25 | 26 | 27 | 28 |
- 인공지능
- GAN
- cs231n
- 기계학습
- 머신러닝
- Deep learning
- python
- 코딩테스트
- 설명가능한
- keras
- Explainable AI
- 설명가능한 인공지능
- 백준
- Class activation map
- 시계열 분석
- Interpretability
- Unsupervised learning
- grad-cam
- meta-learning
- 메타러닝
- AI
- 코딩 테스트
- xai
- 딥러닝
- SmoothGrad
- Machine Learning
- Artificial Intelligence
- Cam
- Score-CAM
- coding test
- Today
- Total
목록Deep learning (20)
iMTE
딥러닝 책 정리를 하였습니다. (읽은 책들 + 읽고 있는 책들 위주로 정리, 극히 주관적이나 책을 사서 공부할 때 도움이 되길 바라면서 정리하였습니다.) 다음 분들에게 도움이 될 것이라고 믿습니다.1. 딥러닝을 체계적으로 code scratch부터 배우고 싶은 분들 (★)2. 이미 잘 구성된 library를 사용해서 모델을 구현하고 싶은 분들 (★★)3. 딥러닝의 이론에 대해서 깊게 공부하고 싶으신 분들(★★★)4. 딥러닝 응용에 대해서 최신 트렌드를 공부하고 싶으신 분들(★★★★)모든 책들은 다 읽어보고 직접 code를 구현을 해보고 결정내린 것입니다. 공부할 때 유익할 것이라고 기대합니다. 1. 추천1) ★, 밑바닥부터 시작하는 딥러닝 (Deep Learning from Scractch) - 딥러닝의..
You Only Look Once : Unified Real-Time Object Detection (2016)Redmon, Joseph, et al. "You only look once: Unified, real-time object detection." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.(학교에서 찍은 영상에 YOLO를 적용한 case. 나무는 class가 없어서 detect을 하지 못하고 대신에 자동차는 잘 detect한다) Object를 단순히 classification하는 문제는, ImageNet, CIFAR 등 dataset으로 학습된 모델 (AlexNet, ResNet, Goo..
Show and Tell: A Neural Image Caption GeneratorVinyals, Oriol, et al. "Show and tell: A neural image caption generator." Proceedings of the IEEE conference on computer vision and pattern recognition. 2015.Image-to-sentence translation은 sentence-to-sentence translation을 기초로, 기존에 Encoder로 사용된 RNN 대신에 CNN을 사용해서 이미지를 encoding을 하였다. Image-to-sentence 작업은 기존의 object detection과 object classification과 달..
starGANStarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image TranslationChoi, Yunjey, et al. "Stargan: Unified generative adversarial networks for multi-domain image-to-image translation." arXiv preprint 1711 (2017). Facial emotion을 detection하는 알고리즘을 만들다가, starGAN 논문을 읽고 정리를 해본다.StarGAN에서 주목한 점은, 기존의 접근 방식들이 (Pix2Pix, CycleGAN, DiscoGAN 등) 두 개 이상의 domain 사이의 관계를 학습할 때 ..
Pix2PixImage-to-Image Translation with Conditional Adversarial NetworksIsola, Phillip, et al. "Image-to-image translation with conditional adversarial networks." arXiv preprint (2017). cycleGAN, DiscoGAN, Pix2Pix 와 같은 image-to-image translation model은 보란듯이 첫 페이지에 결과를 보여준다. (이러니 안궁금할 수가 없지.) 핵심만 간단하게 해석하고 접근을 해보자.먼저, Pix2Pix는 다른 cylceGAN, DiscoGAN과 달리 Paired image를 요구한다. 즉, input output이 서로 관련이 있..
DiscoGANLearning to Discover Cross-Domain Relations with Generative Adversarial NetworksKim, Taeksoo, et al. "Learning to discover cross-domain relations with generative adversarial networks." arXiv preprint arXiv:1703.05192 (2017).GAN에서 직접적으로 결과를 보여줄 수 있는 것은 역시나 이미지인 것 같다. cycleGAN도 그렇고, 이미지로 확 impact를 주니.. 안궁금할 수가 없는 것 같다. cycleGAN과 굉장히 유사한 개념을 기반으로 loss function을 정했고, network를 구성했다. 이 논문에서 cy..
CycleGANUnpaired Image-to-Image Translation using Cycle-Consistent Adversarial NetworksZhu, Jun-Yan, et al. "Unpaired image-to-image translation using cycle-consistent adversarial networks." arXiv preprint (2017).위의 그림에서 보는 것과 같이, cycleGAN은 서로 다른 domain의 이미지를 translate하는 'Image-to-Image translation' GAN이다. GAN이라는 단어가 사용되었기 때문에 당연히, Discriminator와 Generator는 서로 'Adversarial learning'을 시행한다. Cycle..
Super-resolution GAN (SRGAN)Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial NetworkLedig, Christian, et al. "Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network." CVPR. Vol. 2. No. 3. 2017. 간략하게 설명하자면, 기존의 super-resolution 문제에 대해서 단순히 MSE(Mean squared error)를 loss function(objective function)으로 정해놓고 optimization을 해본 결과, high peak sign..
참고 자료:https://www.slideshare.net/ssuser06e0c5/variational-autoencoder-76552518http://jaejunyoo.blogspot.com/2017/04/auto-encoding-variational-bayes-vae-1.htmlhttps://ratsgo.github.io/generative%20model/2018/01/27/VAE/ Variational Auto-Encoder (VAE) Auto-encoder는 high-dimensional data에 대해서 low-dimensional feature를 추출하고 (Encoder) 이 추출된 feature를 기반으로 original data를 복구하는 구조 (Decoder)를 갖고 있다. 개인적으로 참..
Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery Abstract1. Models are typically based on large amount of data with annotated examples of known markers aiming at automating detection.2. Unsupervised learning to identify anomalies in imaging data as candidates for markers.3. AnoGan, a deep convolutional generative adversarial network to learn a manifol..