Lecture Video:
https://www.youtube.com/watch?v=FmpDIaiMIeA&feature=youtu.be&t=1m43s
by Brandon Rohrer
Convolutional Neural Networks(CNN / 卷積神經網路)
CNN: Match pieces of image
Training: Steps of CNN
Get Feature Images
Slice the image into pieces.
Filtering
2.1 Multiply each image pixel by the corresponding feature pixels.
2.2 Add the match result and get average
Convolution
3.1. Try any possible features
3.2. Repeat 1. on other feature. = Convolution Layers.
Pooling (for scaling)
Set window size and get the maximun in the window:
Goal: Get similar pattern, but smaller.
Prediction
Fully Connected Layer
After training by training steps 1-4, we can form the Fully Connected Layer and use it to vote:
Backprop
2.1 Use Backprop to select Parameters of Learning
2.2 Error = right answer - actual answer
2.3 Use Gradient Descent to minimize error
Knobs of CNN
Usage / Limitation of CNN
CNN can be only used for image like problems. (ex. sound, text)
<=> If the solution of the problem is the same after the data column changes, then the problem is not suitable to use CNN
(ex. customer information)
Image: Ok
Sound: Ok
Text: OK
Customer Data : X
