Lecture 5 Convolutional Neural Networks

• Convolutional Network: NN using convolutional layers
  • We want to preserve spatial structure of image! But Fully connected NN cannot do that.
  • preserve (horizon * vertical * channel) structure
  • For 32 * 32 * 3 image,
    • convolve the filter(5 * 5 * 3)(or receptive field for each neuron) with the image spatially
    • elementwise multiplication(dot product), which results in one value
    • with 6 filters, activation map becomes (28 * 28 * 6)
      • Stride is assumed to 1. If assume stride = 3, then the size of activation map becomes (10 * 10 * 6).
        • That is, (32 - 5)/3 + 1 = 10.
        • If stride =2, then (32 -5)/2 + 1 = 14.5 :( ({size of image} - {size of filter})/{stride num} + 1 must be integer!
      • In practice, we tend make zero pad border to make the size of activation map same with image size(when filter size is 3 * 3, zero pad might be 1 pixel. when filter size is 5 * 5, then zero pad might be 2 pixel)
    • include pooling layer(downsampling), which makes representations much smaller to manageable scale
      • Max pooling or Average pooling
  • looking at a local region- spatially! (unlike FCNN- take features from linear vectorization and dot product)

(Omit the history of CNN and range of image recognition & reconstruction)