This summative assignment aims to implement a neural network to classify images in the Fashion MNIST dataset. You will start by using Stochastic Gradient Descent $($SGD$),$ sigmoid activation function and the cross$-$entropy error function. Then, when you have some results, compare with a number of suggested alternatives.

The class in the MNIST dataset are as follows:

$0$T$-$shirt$/$top$1$Trouser$2$Pullover$3$Dress$4$Coat$5$Sandal$6$Shirt$7$Sneaker$8$Bag$9$Ankle boot

Part $1$: Platform preparation, reading data and feature engineering and training

Download or read the dataset here or from other online sources that you know:$$Fashion MNIST $|$ Kaggle$($You will be asked to create an account, we suggest using your Liverpool account$)$

Split the data to train and test. Feel free to use the Keras code facilities or use your coding. You can view the images to obtain a clearer understanding of the data.

Part $2$: Define the model and parameters and start training

Use Stochastic Gradient Descent $($SGD$)$ optimizer, the Sigmoid Activation function, and the cross$-$entropy error function, with a two hidden layer fully connected neural network.

Start training the network with epoch$=10$ and batch size$=1000.$

Obtain the train loss and train accuracy versus test loss and test accuracy.

Part $3$: Repeat Part $2$ implementing the following:

Replace the Sigmoid Activation function with ReLu and examine the results, explaining why they are better or worse.

Use Dropout regularization and briefly explain the changes to the outcomes.