Classify the features of pictures with full connection +softmax

Softmax take logits（ Digital output of the last linear layer of the multiclass classification neural network ） Convert to probability , The method is to get the index of each output , Then normalize each number by the sum of these indexes , So the whole output vector adds up to 1- All the probabilities should add up to 1. The cross entropy loss is usually the loss function of this multi class classification problem .Softmax Often attached to the last layer of image classification network , for example ImageNet Used in competitions CNN（ for example VGG16） Those in .

Softmax Function in CNN Location in Architecture . from neurohive cnn Enhanced images . As shown above ,Softmax The input of is the output of the full connection layer immediately before it , It outputs the final output of the whole neural network . The output is the probability distribution of all label class candidates . source VGG16 — Convolution network for classification and detection https://neurohive.io/en/popular-networks/vgg16/