[Google_Bootcamp_Day16]

Convolutional Neural Network

Computer Vision problems

• Image classification
• Object Detection
• Neural Style Transfer
• etc…

Deep Learning on large images

• Assume image size is 1000 * 1000 * 3, # 1st hidden layer node = 1000
• Then W[1] = (1000, 3M)

Problem : too many parameters
Solution : Use Convolutional Neural Network instead of standard Neural Network

Edge Detection (basic of cnn)

Vertical Edge Detection

Horizontal Edge Detection

key point of cnn : learn “filter matrix” itself

Why padding is necessary

• Image shrink problem
• Edge pixels are used much less

“Valid” and “Same” convoliton

• “Valid” (no padding) : (n * n) * (f * f) -> (n-f+1) * (n-f+1)
• “Same” (pad so that output size = input size) :
  (n * n) * (f * f) -> (n+2p-f+1) * (n+2p-f+1), p = (f-1)/2 (f : usually odd)

Stride

[source] https://www.coursera.org/learn/convolutional-neural-networks