Neural networks have become ubiquitous in our lives, but their inner workings are still baffling even to many practitioners. In this post, I’ll explain how Feedforward Neural Networks conceptually work using just basic linear algebra and geometry.
At their core, these networks learn to partition space of labelled input into regions associated with specific output classes. They do this by applying two key operations: linear transformations and non-linear activations/distortions.
Let’s walk through a simplified example. Say we have 3 classes of labelled data, and our input data has 2 features (X1, X2). We can visualize this as points in 2D space. Our goal is to divide this space so new points with 2D coordinates fall into the correct 3 regions and be classified.