among  m ∈ {0,1}^d  Is to discard the mask （dropout mask）, By taking probability as p The Bayesian effort distribution is randomly generated .

（ The above pictures and concepts come from teacher Qiu's God will network ppt The chapter of regularization ）

In short, it is through incomplete invocation of nodes , Regularize , The purpose is to prevent the model from over fitting

1.5 softmax（ Multiple classifiers ）

          If you use Softmax Regression classifier , Equivalent to the last layer of the network C Neurons , Its output process Softmax The normalized function can be used as the conditional probability of each class .

Is to get the probability of each event output , And then class Determine the category together

2. adopt Matlab Realize a simple convolutional neural network classification

         Here is matlab An example of convolution image classification provided in , Let me explain in detail how each step of the model comes , And explain how we can set it according to our own data set

2.1 Load data

         Load the digital sample data as image data storage .

         The annotated line here adds the file address , We can add ‘c：’ That's what it means c disc , Click the single quotation mark in the back to the middle to select the address .

        imageDatastore Function automatically labels the image according to the folder name .

digitDatasetPath = fullfile(matlabroot,'toolbox','nnet','nndemos', ...
    'nndatasets','DigitDataset');
% digitDatasetPath = fullfile('','','','','');

imds = imageDatastore(digitDatasetPath, ...
    'IncludeSubfolders',true, ...
    'LabelSource','foldernames');   

         Open the data set, which is like this , The file name is the label , Inside is the data of handwritten digits

         The label here doesn't have to be a number , For example, when I was doing it, there was the English of the corresponding fault classification name , I haven't tried Chinese , I don't think most of them support .

2.2 Divide the data set

         Divide the data into training data set and verification data set , So that each category in the training set contains 750 Images , And the validation set contains the remaining images corresponding to each tag .splitEachLabel Split the image data store into two new data stores for training and verification .

         here 750 Is the number of training sets , Here we need to pay attention to , This number should not be greater than the minimum number in each folder , Otherwise, it will report a mistake , hinder randomize It's just to take randomly , It's convenient that you don't need to disturb manually .

numTrainFiles = 750;
[imdsTrain,imdsValidation] = splitEachLabel(imds,numTrainFiles,'randomize');

2.3 Define the network framework （ It is suggested to watch the video at the top of the article ）

         Define convolutional neural network architecture . Specify the size of the image in the network input layer and the number of classes in the full connection layer in front of the classification layer . Each image is 28×28×1 Pixels , Yes 10 Classes .

inputSize = [28 28 1];
numClasses = 10;

layers = [
    imageInputLayer(inputSize)
    convolution2dLayer(5,20)
    batchNormalizationLayer
    reluLayer
    fullyConnectedLayer(numClasses)
    softmaxLayer
    classificationLayer];