Inaccuracy of traditional neural networks when images are translated

To understand the need of CNNs further, we will first understand why a feed forward Neural Network (NN) does not work when an image is translated and then see how the CNN improves upon traditional feed forward NN.

Let's go through the following scenario:

• We will build a NN model to predict labels from the MNIST dataset
• We will consider all images that have a label of 1 and take an average of all of them (generating an average of 1 image)
• We will predict the label of the average 1 image that we have generated in the previous step using traditional NN
• We will translate the average 1 image by 1 pixel to the left or right
• We will make a prediction of the translated image using our traditional NN model