1. 池化层
: Extract important information, remove unimportant information, reduce parameters, reduce computational overhead, and prevent overfitting.
2. 全连接层(FC)
: It acts as a "classifier" in the entire convolutional neural network.
3. 激活函数
: Introduce nonlinear factors to neurons, thereby improving the expressive ability of the network.
4. backbone
: 主干(骨干)网络
, generally refers to the network that extracts features, and its function is to extract the information in the picture for use by the subsequent network.
5. 反向传播
: In backpropagation, the movement of the network is backward, the error flows in from the outer layer with the gradient, passes through the hidden layer, and the weight is updated.
6. 超参数
: It is a parameter whose value is set before starting the learning process, not the parameter data obtained through training.
7. normalization:
Data standardization processing to solve the comparability between data indicators.
Batch Norm
On the batch, normalize the input of each single channel;
Layer Norm
in the direction of the channel, normalize the input of each depth.
8. MLP
: The Chinese name is 多层感知机
, its essence is the neural network. It is proposed mainly to solve nonlinear problems that single-layer perceptrons cannot solve. Detailed blog introduction
transformer