Some methods of himself analysis (II.) will be supplemented in the future. --by wepon
Combined with the literature "deep Learning for computer Vision", here are some points of attention and questions about convolutional neural networks.
The excitation function is to choose a nonlinear function, such as tang,sigmoid,rectified liner. In CNN, Relu is used more because: (1) Simplifying BP calculations and (2) making learning faster. (3) Avoid saturation problem (saturation issues)
Pooling (Pooling): Its role in (1) to some small morphological changes remain invariant, invariance to small Transformations, (2) have a greater sense of the domain, Larger receptive fields. The pooling method has sum or max.
Normalization:equalizes the features maps. It has the following functions: (1) introduces local competition between features; (2) Also helps to scale activations on each layer better for Learni ng; (3) empirically, seems to help a bit (1–2%) on ImageNet
Training Cnn:back-propagation;stochastic Gradient descent;momentum;classification LOSS,CROSS-ENTROPY;GPU implementation.
Pretreatment: Mean removal;whitening (ZCA)
Enhanced generalization capability: Data augmentation;weight regularization; adding noise to the network, including dropout,dropconnect,stochastic pooling.
- Dropout: The output of some neurons in the fully connected layer is randomly set to 0 at the full connection layer only.
- Dropconnect: Also only used on the full-connection layer, Random binary mask on weights.
- Stochastic Pooling: Convolution layer used. Sample location from Multinomial.
Model is not work, how to do? In my own experience, the initial value of the learning rate is set too large to start with 0.01, which is small, but actually 0.001 is more appropriate.
Some details of convolutional neural networks