Hidden Layers Learn Intermediate Features

Compose small calculations

A layer applies a group of neurons to an input vector. The input layer holds raw features; one or more hidden layers create intermediate representations; the output layer produces the task result. Hidden means those values are internal, not that they are mysterious or encrypted.

pixels -> edges -> shapes -> object score
tokens -> local relations -> contextual meaning -> next-token scores

Early layers can represent simpler reusable patterns, while later layers combine them into task-specific concepts. This hierarchy is learned from the loss rather than assigned names by programmers. Not every neuron maps cleanly to a human concept, so the diagram is intuition, not a guaranteed literal decomposition.

Analogy: A factory line first turns raw material into parts, then parts into assemblies, then assemblies into a product. Each station consumes the representation produced by the previous one.

Note: Deep learning simply means several representation-learning layers are stacked. Depth is valuable only when the task, data, and compute justify the added complexity.