Article directory
The fourth chapter deepens the network and improves the performance of the model
4.1 Why a Deeper Network is Necessary
(1) Stronger expressive ability
The main modules of the current deep learning model structure are convolution, pooling, and activation , which are standard nonlinear transformation modules.
But the model is more used in the combination of convolution, Batch_norm, Relu, and Dropout .
(2) Learning is easier.
CNN abstracts information through layers. As the network deepens, the level of abstraction also deepens. Deeper network layers can learn more complex representations. A shallow network means that the expressions to be learned are very complex, which is difficult to do. If the network is deep enough, the transformation that each layer needs to learn is relatively simple, so that the model performance is better.
4.2 Network structure diagram of VGG16
Carefully observe the model design:
(1) pool design
(2) dropout design
4.3 Selection of optimizer and learning rate adjustment strategy
In this book, the conclusion is obtained through experiments:
In the classification experiment, use:
(1) SGD (momentum m=0.9)
(2) Better learning rate adjustment strategies are: Step, Multistep