Deploy your apps to a supercloud in a few clicks

This Engineering Education program is supported by Section. Instantly deploy your GitHub apps, Docker containers or K8s namespaces to a supercloud.

Try It For Free

Getting Started with Deep Neural Networks in Matlab

May 10, 2021

In this article, we will look at how to prepare your digit dataset, the structure of a convolution neural network. We will also look at the Matlab code for training and validation and code for discrete testing.

Creating deep neural network in MATLAB

A neural network is a subclass of machine learning. The algorithms used are a result of inspiration from the architecture of the human brain. The system grips data then uses the algorithm to identify the trend in the data and predicts the result of a new similar dataset.


MATLAB provides a deep learning toolbox for implementing the deep neural network. MATLAB includes the digit dataset so you don’t need to download it from an external source.


To follow this article along - you will need the following:

  1. Knowledge in Matlab.
  2. Matlab installed on your computer.

In this tutorial, we will tackle the problem of digit recognition.

MATLAB’S digit dataset

Matlab itself includes the digit dataset. The digit dataset consists of 10,000 grayscale images of handwritten digits and each digit has 1000 images of the respective folders. The images are of size 28 by 28 pixels and they have associated labels.

How to locate the image dataset

  • Open the Matlab installation directory.

this is how the window appear

  • Select the toolbox folder.


  • Click the nnet and then choose the nndemos folder.



  • Select the mdataset folder.


  • Select the digit database folder.


When we open this folder, we get ten folders that contain 1000 images of each number.

this how the digit database appears

When we open a folder of any number we will be able to see the 1000 images.

Below is sample images in folder 1.

images in folder 1

Note that after locating the dataset folder, you can duplicate it to your preferred folder.

For training, we will take 750 images from each folder of the digit dataset. This means that the total training images will be 7500 images since there are 10 folders. The remaining 250 images sum up to 2500 images will be used for validation. This division is done by writing Matlab code.

Structure of convolutional neural network (CNN/convnet)

CNN is the common category belonging to neural networks. CNN revolves around the features discovered from the input data. The network operates a 2D convolutional layer placing this constitution in a better position to analyze data for instance images. The first layer of the CNN is the input layer that we have to define when we will define the CNN in Matlab.

We then define the various layers during feature detection. The CNN is made up of 3 layers. The top layer is the input layer. The middle layer includes a 2D convolutional layer, batch normalization layer, relu layer, max pooling layer. The last layer involves a fully connected layer, softmax layer, and classification layer.

The second layer which has 4 layers will be used repeatedly. (There is not only four because we have different bunches in those 4 layers.)

Matlab syntax of the layers and the specifications

Input layer

This is where we specify the image size.

ImageInputLayer([M N n], ‘name’, ‘input’)


ImageInputLayer([28 28 1], ‘name’, ‘input’)

The M and N depend on the size of the image. The ‘n’ is the no. of layers of the images. In our example, we used 1 since the grayscale is a single layer. If your image is colored, then ‘n’ will hold the value of 3, since colored images have 3 layers, that is, RGB.

Convolutional layer

The main layer is responsible for the features extraction. Filters keep updating with training data.

convolution2dLayer(filterSize, numFilter, ‘stride’, n, ‘padding’,’same’,’name’,’conv_1’)
convolution2dLayer(3,8,’padding’, ‘same’,’name’,’conv_1’)

In this example we will have a 3X3 pixel for filter size and they will be 8 filters. The no. of filters means the no. of neurons in this layer while the strides mean by what pixels should the filters shift to the next group of filters.

Its default value is one but it varies according to your precision. In padding, you also require to cover the first column or first group. Adding ‘same’ means after doing the padding, the resulting image, and the input image should be of equal size.

Batch normalization layer

This normalizes the activation and gradients. This eases the optimization problem, quickens the training of the network, and minimizes the sensitivity to initializing the network.


Relu layer (rectified linear unit)

This is the popular random activation function:


Max pooling layer

The max pooling layer is an operation of downsampling the features in an image to predict the future image.


Fully connected layer

The fully connected layer analyzes sampled features and categorizes the images.


The output represents the no. of neurons in this layer. In my example, we used 10 since the output size depends on the categories of objects diagnosed using the CNN.

Softmax layer

THis has an activation function that changes the result of the fully connected layer. The result of the softmax layer are positive numbers that add to one and might be used by the classification layer for classification possibilities.

softmaxLayer(‘name’, ’name’)

Classification layer

This is the final layer. It solves the possibilities from the softmax activation function to their predicted form.

classificationLayer(‘name’, ’name’)



Training parameters


In the training option, we have to define the solver name.



We used the plot factor here to see the progress of the training in a graph. The graph will be of output against epochs. There are various parameters and if you want to learn all, check the documentation.

Solver options

  • sgdm: stochastic gradient descent with momentum optimizer. It needs a momentum rate.
  • rmsprop: RMSProp optimizer. The squared gradient moving average decay rate must be included.
  • adam: adam optimizer. The gradient and squared gradient moving average decay rate must be included.

Matlab code for training and validation

We first give the path of the dataset folder and read the image database folder.

digitDatasetPath = ‘directory’;
digitImages =    imageDatastore(digitDatabasePath,’includeSubfolder’,true,’labelSource’,’folderNames’

We then distribute the images in the set of training and testing using the code below:

numTrainingFiles = 750; %you can use percentages to represent the divisions.    
[trainImages,testImages] = splitEachLabel(digitImage, numTrainFiles,’randomize’); %the spliting is done randomly.

Building of the CNN

We first define the different layer of the CNN:

layer = [
    imageInputLayer([28 28 1],’name’,’input’)

Note that the CNN layer is repeated but you have to increase the size of the filters. This gives more complexity to your network and more strength and capability to your network.


In this last part of the CNN we don’t use the max pooling layer.


This is the general structure of the network. If you want to see this network graphically, you have to execute the code below.

Igraph = layerGraph(layers)

Training option

This is the last section:

option = trainingOptions(‘sgdm’,’initialLearnRate’,0.001,’maxepochs’,4,’shuffle’,’everyepoch’,   ‘validation data’,’test images’,’validation frequency’, 30, ‘verbose’,’false’,  ‘plots’,’trainingProgress’)

We then initiate the training:

net = trainNetwork(trainImages,layers,options);

We then classify the images:

Ypred = classify(net,testImages)

Once the training is done, you can export this function net to be used by someone else to identify the various objects. We then run the script in Matlab, this may take some time since there is a lot of images being used. The first graph is for training and the second graph is for the loss of error in the workspace.

graph for training process

Matlab code of discrete testing

We create a separate script for discrete testing and call the training option in it.

We first read the image for clasification:

[fileName, pathName] = uigetfile(‘*.*’,’select the input grayscale’);
fileWithPath = strcat(pathName,fileName)
I = imread(fileWithPath);

We then classify the image using the network.

Label = classify(net,I);
title([‘recognized digit is:’ char(label)])

recognized dataset

Apart from the images in the dataset, you can create your image and recognize it using the program. This image should be a grayscale image.


Matlab is a good language for deep neural systems. It has a toolbox that provides data that can be used for the training and the testing of the neural system.

Since it has the dataset to be used for the process, this makes it more efficient for use.

Happy coding!

Peer Review Contributions by: Odhiambo Paul