Why this matters
Neural networks sit behind speech recognition, image models and language models. You will rarely design a big one from scratch in your first roles, but you must understand what layers, activations, epochs and loss mean, or every training log and fine tuning guide will look like noise. Ghanaian AI adverts name PyTorch, TensorFlow and Keras for this reason.
One unit, then a layer
A single unit (often called a neuron) does what you already know from the maths section: a dot product of inputs and weights, plus a bias. Then it applies an activation function. The most common is ReLU, which keeps positive numbers and turns negatives into zero.
A layer is many units side by side, each with its own weights. In matrix terms, a layer is one matrix multiplication followed by the activation.
Why the activation matters
Without activations, stacking layers is pointless: a chain of straight line calculations is still a straight line. The activation adds a bend. With enough units and layers, those bends let a network fit very complicated patterns.
Training
Training a network is gradient descent, the same loop from the maths section:
- Forward pass: feed inputs through the layers to get predictions.
- Loss: measure how wrong they are.