Stanford-university (cs231n)

• Idea:

  • Dig deeper into convolutional neural networks (CNN)
  • Implement each CNN parts from scrach without NN libraries
  • Finish one of the greatest's university in the world (Stanforn university) semester on computer vission - cs231n

• Environment:

  • Atom
  • Jupyter notebook
  • Python 3

• Datasets:

  • CIFAR-10
  • COCO and ImageNet

• Categories:

  • Assignment 1

    • KNN (k-Nearest Neighbors)
    • SVM (Support Vector Machine)
    • Softmax
    • Two-Layer Neural Network
    • Higher Level Representations: Image Features

  • Assignment 2

    • Fully-connected Neural Network
    • Batch Normalization
    • Dropout
    • CNN (Convolutional Neural Networks)
    • TensorFlow on CIFAR-10

  • Assignemnt 3

    • Image Captioning with Vanilla RNNs (Recurrent Neural Networks)
    • Image Captioning with LSTMs (Long short-term memory)
    • Network Visualization: Saliency maps, Class Visualization, and Fooling Images
    • Style Transfer
    • GANs (Generative Adversarial Networks)

• Some outputs

  • Assignment 1

    • KNN accuracy
      
    • SVM weights visualization
      
    • 2 layer NN training process
      

  • Assignment 2

    • NN training with different optimizers
      
    • Model sensitivity with Batch normalization
      
    • Model sensitivity with Batch normalization
      
    • Model accuracy with Dropout
    • CNN filters visualization

  • Assignment 3

    • Training image RNN
      
    • Training image LSTM
      
    • Style transfer
      
    • GANs output