day18.md

Day18 (2022/02/09)

노트

Core CNN Models

Baselines

• AlexNet
  • Used ReLU activation
  • Local Response Normalization
  • Overlapping Pooling
  • Data Augmentation
  • Dropout
• VGGNet: Used only 3x3 convolutions
• GoogLeNet: Used wise 1x1 convolutions
• ResNet: Residual Network
• DenseNet: Concatenation Network

Semantic Segmentation

• Fully Convolutional Network
• UNet: Auto-encoder Network
• DeepLab

Object Detection

• R-CNN
  • Selective Search
  • Bounding Box Regression
  • SVM
• SPPNet: Spatial Pyramid Pooling
• Fast R-CNN: ROI Pooling
• Faster R-CNN: Region Proposal Network
• YOLO
  • Without explicit bounding box sampling
  • Simultaneous prediction of bounding boxes and class probabilities

Generative Models

• Implicit Models: Generation only
• Explicit Models: Estimate density

Related Tasks

• Generation
• Density Estimation
• Unsupervised Representation Learning

Auto-regressive Models

• Leveraging conditional independency
  • Neural Autoregressive Density Estimator: Consider every prior pixel
  • Pixel RNN: Consider some prior pixels
• Ordering data patches
  • Pixel RNN with Row LSTM: Consider every top-left prior pixel
  • Pixel RNN with Diagonal BiLSTM: Consider diagonal prior pixels

일지

Daily scrum (10:00-10:10)

강의 영상 수강 및 퀴즈 제출 (10:10-12:00)

• [강의] Deep Learning Basics
  • Modern CNN
  • Computer Vision Applications
• [퀴즈] Deep Learning Basics
  • CNN

강의 영상 수강 및 퀴즈 제출 (13:00-14:30)

• [강의] Deep Learning Basics
  • Generative Models I

과제 수행 (14:30-16:00)

• [과제] Deep Learning Basics
  • ViT

Peer session (16:00-17:00)

• Transformer 관련 이야기
• 대학원 연구실 관련 이야기

Daily report 작성 (17:00-19:00)