Page 41 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

Dinh-Hoang-Long Tran, Quoc-Huy Le 25

Fig. 2. Structure of Convolutional Neural Network (Source: Developersbreach [10])

Convolutional Layer which its purpose is to extract features of the input layer. In the
layer, Kernel behaves like a filter that extract the features by moving along the matrix
of input layer. It has the same dept of the input layer the value of kernel called weight.
The Kernel moving along to all value of input layer by step, the magnitude of that step
is Stride.

Pooling layer: The purpose of this layer is reducing the size of date by decrease the
dimension of previous convolution layer but keep importance features.

Fully connected layer: After going through countless Convolutional Layers and
Pooling Layers to get the optimal feature data, the most important is the Fully
Connected Layer is the group of layers that determines the outcome of the entire model.
All the feature of the input will convert to 1 dimension vector through the Flatten layer
for ready to put in activation function for final output.

Activation function: Make the result nonlinear. The combination of activation
functions between hidden layers is to help the model learn nonlinear relationships and
hidden problems in the data (Fig. 3).

Fig. 3. Activation function

ISBN: 978-604-80-8083-9 CITA 2023