Digit Characterisation with Neural Networks23 Jun 2016
This post is part of the series Neural Networks.
A common application of neural networks is pattern recognition in images. Since we humans can effortlessly recognise shapes such as characters in text, it seems surprising at first that programming a computer to do the same thing is so difficult. Writing an algorithm to extract and analyse the shape of a handwritten digit quickly becomes extremely complex as all the possible permutations of shape and orientation, colour and intensity of the digit and its background, etc. must be considered.
Using a neural network, on the other hand, allows us to leverage on the computer similar processes to what is going on in our brain. The network, which knows no specific information about the task to be performed, can be trained to a set of inputs and outputs, with the training algorithm automatically strengthening the right connections. The implementation and training of neural networks was covered in detail in the previous posts of this series.
In this post, I will introduce the application of neural networks to the characterisation of handwritten digits. My neural network implementation and all the code required to reproduce the results presented here are available on GitHub.
For this application of my neural network implementation I used the excellent MNIST database, which contrains a total of 70,000 images of handwritten digits and their corresponding labels. The images and labels are provided in a compressed binary format, for which I have provided an interpreter in my GitHub repository.
Using this tool you can easily download, read and display a random sample of the MNIST images:
This will look something like:
While the MNIST data set comes with 60,000 training images and 10,000 test images, we will reserve 10,000 images from the training set for cross validation. With any machine learning algorithm, it is important to retain a cross validation data set to compare the efficacy of different methods or learning parameters that is independent of the test set.
In the MNIST data set, each image is a 28x28 image of grayscale intensity values. Since each pixel value constitues a single input feature, the first layer of the neural network will be 784 in size. The output layer, which will consist of one activation for each possible digit, will be 10 in size. The number and size of the intermediate layers of the network will be a design choice made by us.
For starters, let’s consider a network with a single hidden layer of size 300, giving a network with layers of sizes 784, 300, 10. Using my Python neural network code, this is created as follows:
Training the Network
To train the network, the mnist.py module referenced above is used to first load the training data and labels:
These are then randomly shuffled and divided into training and cross validation sets (the separate testing dataset files will be used for final testing of the network).
Finally, the image arrays are reshaped so that each pixel value is a single feature:
These feature and label arrays can be used to train the neural network:
and then make predictions with the
train method of the neural network takes an optional argument allowing
the tracking of the accuracy of the network on the cross validation dataset over
the training process. To get the
train method to return the history of the
accuracy (fraction of correct predictions) of both the training and cross
validation datasets, run it as follows:
We can then plot the history of the accuracy of the iterations of the training algorithm:
Yielding the following:
Here we can see that the training algorithm starts with an accuracy of around 0.1, or 10%, which is what you would expect for getting the answer right by accident 1/10 of the time with the randomly initialised parameters. The accuracy of both the training and cross validation datasets improves rapidly, before stabilising with a prediction accuracy above 93% after around 50 iterations. After this, the method appears to have converged, with only modest gains in accuracy achieved.
Importantly, the fact that we get good accuracy with the cross validation dataset indicates that our network is not being overfitted to the training set, and so we will have confidence that it can be applied to similar new datasets with confidence. In later posts in this series, we will modify the network and training algorithm to both optimise the accuracy of the network and make the training process as efficient as possible.