The URL of: http://ydwen.github.io/papers/WenECCV16.pdf
The main contribution of this paper is to propose a Center Loss of function loss, and the use of Softmax Loss Center Loss to oversee joint training, while expanding the class differences between abbreviations class differences within and enhance the robustness of the model.
To illustrate the effect of visual softmax loss, the author made a simple change of LeNet, the output of the last hidden layer dimension to 2, characterized in that the two-dimensional plane and the visual, the following two pictures are of the train set and test MNIDST set, the differences between classes can be found in more obvious, but the difference in the class also obvious. 
In order to reduce class differences within the paper proposes Center Loss: Big Box Center Loss - A Discriminative Learning Approach for the Feature Deep Face Recognition -Deep-Face-Recognition-image004.png "alt =" "/>
C yi central point is the class wherein, the mean is calculated Cyi yi type of sample characteristics, in order to allow center loss practical neural network training process, C yi calculated for each mini-batch concerned, thus binding Softmax loss, loss of the entire network becomes a function, [lambda] to balance these two Loss: 
simply replaced by the same circuit construction as Softmax Loss Center Loss of data sets on MNIST did the same experiment for different [lambda] is worth to visualize the result as follows Center Loss can be found quite significantly reduces the within-class differences while differences between the classes is also outstanding. 
performance data set in the open: