Hidden layer

Think of a mortgage-approval model. The input layer receives raw numbers such as income, debt and credit history. The output layer gives a single approval score. In between, one or more hidden layers combine those numbers into intermediate signals, such as something like “debt is high relative to income”, that nobody wrote down as a rule. Each hidden layer is a set of artificial neurons that take the previous layer’s outputs as their inputs.

Why “hidden”. The training data tells the network what the input is and what the right output should be. It never says what the layers in the middle should compute. The Deep Learning textbook uses that as the definition: these layers are hidden because the data does not show their desired output. Training with backpropagation decides what they learn.

Depth. A multilayer perceptron has at least one hidden layer. “Deep” in deep learning means many hidden layers stacked on top of each other. A transformer language model is also built from a stack of repeated blocks whose intermediate outputs are hidden states. Stacking more layers lets a model represent more complex patterns. It also makes the model harder to train, which is one reason residual connections were invented.

Why it matters. Hidden layers are where a neural network’s useful behavior comes from, and also where it becomes hard to explain. A spreadsheet rule can be read line by line. The features inside a hidden layer are lists of numbers, so explaining a model’s decision means studying those numbers indirectly. That gap is the practical reason businesses ask for model documentation, evaluation and monitoring rather than relying on reading the model itself.