Trace a tiny model that takes a few input numbers, multiplies them by weights, adds a bias, and produces a first guess. You’ll see each weight as a knob that makes one input matter more or less.

Follow how hidden layers pass signals forward and reshape simple inputs into more useful intermediate patterns. You’ll recognize why activations such as ReLU let stacked layers learn more than one straight-line rule.

Apply the previous explanations in a guided problem.

Compare how the last layer represents different prediction tasks: a number for regression, a yes/no score with sigmoid, or class probabilities with softmax. You’ll decide what the model’s raw output means before calling it a prediction.

Calculate how far a model’s guess is from the right answer using a loss. You’ll compare mean squared error for number guesses with cross-entropy for classification and see why training needs one score to improve.

Apply the previous explanations in a guided problem.

Check your understanding with a short quiz.

Reason through how a model changes its weights after an error using gradients, backpropagation, and a learning rate. You’ll see gradient descent as a small step that nudges the model toward lower loss.

Walk through the repeating cycle of training: make predictions, compute loss, update weights, and try again over mini-batches and epochs. You’ll interpret falling, flat, or unstable loss as clues about whether learning is working.

Apply the previous explanations in a guided problem.

Check your understanding with a short quiz.

Review this chapter with practice based on your mistakes.