DL : Basic Concept of CNN

Part 4.1 of Deep Learning Specialization

CNN structure

1. Convolutional Layer

2. Pooling

3. Flattening + Fully-Connected Neural Network

1. Convolutional Layer

Filter in image processing — Feature Extraction

- weight

- padding

- stride

2. Pooling layer

- Max Pooling

- Average Pooling

** Pooling layer — No parameter to learn **

3. Flattening + Fully-Connected NN

convert pooling layer to input layer of Fully-Connected NN

Why CNN, not NN ?

** Parameter of CNN is less than NN **

input layer 3072 node >> activation node 4704 node

CNN : convolution layer with 6 filter of (5*5 + 1 bias )

= 156 parameters

NN : (3072 input node)* (4704 activation node) + 1 bias

= 14,450,688 parameters !!

Image Augmentation

Increase number of training data

Increase variation of training data

Transfer Learning

- Not Learning new tasks from scratch

- Use previous knowledge to help learning new tasks

- Learning process can be faster & need less training data

Related Articles

• Neural Network — Part 1.1
• Neural Network — Part 1.2
• Neural Network — Part 2
• Neural Network — Part 3
• CNN — Part 4.1
• CNN — Part 4.2
• RNN — Part 5.1
• RNN — Part 5.2

Reference

Deep Learning Specialization: Convolutional Neural Network (Coursera) (Youtube)