-unfinished-

ML-MNIST-Dataset

Using the MNIST dataset to develop a machine learning algorithm that can "read" numbers. The aim of this project was to learn make a clean Convolutional Network Machine Learning Algorithm in Java. To fully understand how a neural network functions, I did not want to depend on modules that permitted a comfortable level of abstraction like the ones in python using numpy. This allowed me to manipulate the matricies directly involved in the network as well as the back propogation.

Additionally, a very important feature of the code was its adaptability with respect to the inputs:

• Kernel size
• Stride
• Pooling method
• Number of layers
• Number of filters
• Number of fully connected nodes
• etc This was, once again, to be able to see what changes affected the efficiency of the code and to what extent. The eventual goal was to devise a method to automatically optimize the value of these parameters to optimize the analysis.

Although this was my first try with neural networks, working with a convolutional network also allowed to work with the back propogation of fully-connected layers Benefints of CNN -> for large images, less nodes?

Implementation

Set up

The entire algorithm is controlled from the Main function in MNIST_Dataset/Files/Main.java. First the set and network are initialized:

• set : loading of all the images (either TRAINING or TESTING)

MNISTSet set = MNIST.load(TRAINING);

• network : contains the hidden layers

MNISTNetwork network = new MNISTNetwork(784, 100, 50, 10); The first layers correspond to the number of convolutive filters are used per stage, and how many stages there are. Then the number of fully connected neurons are passed as arguments with the last entry representing the number of dense layers. Most of the parameters can be changed with the constants file.

NEED TO BE CAREFUL OF FUNCTIONS THAT RETURN A PARAMETER. THEY DO WEIRD THINGS IN JAVA

Training : Convolution

Training : Pooling

Training : Fully Connected Layer

Training : Backward Propagation

of maxpooling : https://datascience.stackexchange.com/a/14235

Testing of Network

Visualisation of Network

As the aim of this project was to understand the steps needed to be taken for a neural network. And the benefits of using a convoluted neural network is from the spacial/feature recognition. Seeing the result of the various convolutive layers is interesting, and to see how they evolved. []visualizing CNN