Deep Learning CNN

CNN

• Basics
  • Tensors
  • Data whose dimension ≥ 3
  • For image data, input tensor = width x height x numberofchannels
  • output tensor = width x height x depth
  • Convolution
  • Specific features (patterns) of an image is only in a local area of the whole image
  • Goal is to get a scanner (filter / kernel) that can detect those features, if there is a match wherever on the whole image, it reacts
• Convolutional layer
  • Network structure
  • layers have neurons arranged in 3 dimensions: width, height, depth
  • depth slice:
    • for one unit of depth, the slice of it is called depth slice
    • all neurons in each depth slice use absolutely the same weight and bias
    • weights are structured in a matrix, which is called kernel
    • each slice has its own kernel
    • For example in p11:
    • 3×3 grid is all the weights being used in that 6×6 slice
    • the value in that 4×4 is called scaler, indicates the reaction level of the filter, the higher the more intense
  • Sliding / Convolving
  • The process of receiving 6×6 as input feature, slide that 3×3 filter, and get the 4×4 output feature
  • For all slices, they always look at the same region, and each slice use a different filter in convolving process
  • Layer depth, Stride and zero-padding
  • layer depth = # of filters
  • stride = how much a filter moves per step
    • stride = 1 -> move one pixel per step
  • 0-padding -> pad the image with 1 or 2 pixel-wide ring of 0
    • used when features are near the boundary
  • number of params see P20
  • that 10 thing is number of channels, for RGB it’s 3, etc
• Pooling layer
  • wtf is this:
  • Subsampling the image to make the image smaller while retaining the object itelf
  • Bring down number of parameters and avoid overfitting
  • how to do this:
  • max pooling, see P22
    • move the window, only output the max value in that window
  • and other fucking pooling shits
• When close to output payer, the tensor is flattened, and fed through FC layers to get the final output
• In training process, still use that fucking gradient descent shit, only CONV and FC are trained because they have params
• Transfer learning:
  • Use a pretrained model, remove the last FC layer, train a new FC layer and connect that layer
• 29-33, 35-40 use cases
  • ECG classes
  • Medical image classification
  • Fish recognition
  • Image style transfer
  • Image sentiment analysis
  • Image recommender

TensorFlow