MNIST classification is the 'Hello World' of deep learning.
The model is a feed forward (Sequential in keras) deep neural network with 2 hidden layers.
Input layer - flattening layer which converts the multidimensional input dataset (in our case, 2D dataset (28x28)) into a single dimensional vector.
Hidden layers - Dense layers with 128 units (neurons) with ReLU (Rectified Linear Unit) activation.
Output layer - Dense layer with 10 units (since our data has 10 classes) with softmax activation.
Model is trained for 3 epochs with adam optimizer and sparse categorical crossentropy to calculate loss.