4. ■Previous neural network
Fully-connected : Affine layer
• Combines with all neurons in adjacent layers
Intro
Interaction Lab., Kumoh National Institue of Technology 4
5. ■Problem of previous neural network
Data is ignored
• Image consists of width, height, and channel 3D information
• Flatten 3-D data into 1-D data
Intro
Interaction Lab., Kumoh National Institue of Technology 5
6. ■Convolution Neural Network
Plus Convolution layer and Pooling Layer
• Use 3-D data
■ Advantage of data shape
Intro
Interaction Lab., Kumoh National Institue of Technology 6
9. ■Padding
Fills around the data with specific values
Resize the output to large
Convolution Layer
Interaction Lab., Kumoh National Institue of Technology 9
(4, 4)
Padding : 1
(3, 3)
Filter
(4, 4)
Output
10. ■Stride
Stride : 2
Resize the output to small
Convolution Layer
Interaction Lab., Kumoh National Institue of Technology 10
12. ■3-D data convolution
Plus channel
• Input channel = filter channel
Convolution Layer
Interaction Lab., Kumoh National Institue of Technology 12
13. ■3-D data convolution
Output with multiple channels
• Use multiple filters
Convolution Layer
Interaction Lab., Kumoh National Institue of Technology 13
14. ■Batch
3-D data -> 4-D data
• (H, W, C) -> (Data, H, W, C)
Convolution operation on n data
Convolution Layer
Interaction Lab., Kumoh National Institue of Technology 14
15. ■Pooling layer
Not use all data
• Pooling window = size of stride
■ 2 x 2 window 2 stride
■ 3 x 3 window 3 stride
■ 4 x 4 window 4 stride
Pooling Layer
Interaction Lab., Kumoh National Institue of Technology 15
2 x 2
window
16. ■Pooling layer
No parameter for training
Not change channel size
Stable
Pooling Layer
Interaction Lab., Kumoh National Institue of Technology 16
17. ■Im2col
Not use for loop
• Performance : numpy
Implement
Interaction Lab., Kumoh National Institue of Technology 17
input
4-D data to 2-D data
Filter application order