CNN for Classification

We saw in the previous lesson that a multilayer perceptron can be trained to classify tabular data. An image differs from tabular data in the sense that an image contains unstructured information. We cannot go to a predefined pixel to extract a useful feature for the classification. Objects in an image dataset can appear anywhere under various poses. For this reason, we generally cannot treat an image as a vector of length $H\times W$ and process it with a multilayer perceptron.

A better approach is to extract high-level features through a composition of convolution layers and spatial pooling. At some point, the spatial resolution is sufficiently low, and we can flatten the image into a vector. This vector can then be processed by a multilayer perceptron.

Problem statement

In this lesson, our task is to build a CNN with the building blocks we studied in the previous lesson. The task is to classify monochrome images of handwritten digits from the MNIST dataset. Each image has a size of 28x28. The classes are the 10 digits, from zero to nine.