Convolutional Neural Network

 

https://towardsdatascience.com/convolutional-neural-network-1368ee2998d3

 

The portion of the image is stored in a single cell, after the operation of the filter

Cell is regarded as neurons, when combined, forma 2D matrix called Activation Map/ Feature Map

The neighborhood which is pointed by a single neuron is called a Local Receptive Field.

 

While training the network, the weights between the neurons (weight of filter) act as a parameter that is tweaked while training

---

---

- Param# of conv2d_2 = weight + bias = 3*3*32 + 32 (if kernel size is 3*3)

- Param# of conv2d_3 = weight + bias = 3*3*32*64 + 64 (if kernel size is 3*3)

- input * weight + bias라고 생각하면 편하다. FCN에서 bias의 수는 output dim의 수, CNN에서는 output channel의 수 이다.