Please enable JavaScript.

Coggle requires JavaScript to display documents.

Convolutional Neural Networks, Invariant a function, quantity, or property…

- - - - Negative Values (<0 or <<0) activate dark pixels.
      - Neutral values (0) doesn't activate anything
      - By combining positive and negative values you can detect features as edges (where the difference in rows or columns between positive and negative indicates the presence of edges because an edge is a region where the pixels differ)
      - Positive values (>1 or >>2) activates bright pixels
    - - Same Convolution: padding allows us to preserve the same height and width volumes after each convolution. Additionally, it helps us keep more of the information at the border of an image.
      - Valid Convolution: shrink the height and width volumes after each convolution.
    - - The 2D output of the convolution is called feature map
      - A convolutional layer is where the filter slides (using a stride) over the image and computes the dot product
      - Extracts features from an input image by taking the dot product between the input data and a 3D array of weights (the filter).
- - - - VGG includes
        
        Block of same convolutions (to focus less in fine-tuning)
        
        Simplify the architectures
      - LeNet: Basic architecture built using
        
        valid convolution
        
        average pooling
        
        sigmoid/tanh
        
        convolution per channel
        
        non-linearity after pooling
      - AlexNet: includes
        
        ReLU activations
        
        dataset augmentation treatment known as PCA color shifting
        
        Local Response Normalization (LRN) (not used now)
  - - - Depends on the training data size
      - Small data: training only the top layer
      - Large data: unfroze some last layers (train them)
      - Huge data: use as initializer and train the whole net
    - - More Used
        
        Mirroring (Flip Vertical)
        
        Random Cropping
        
        Color Shifting
      - Less Common
        
        Rotation
        
        Shearing
        
        Local warping
      - You can apply distortion into a batch in parallel (using threads) while processing the training
    - - It seems like although there are huge dataset nowadays in CV, there is a necessity for more data yet (where it takes place more hand-engineering than other task)
      - Benchmark and Competition Tips: Ensembling and Multi-crop at test time