Summary and Revision Notes

Key ideas

  • A model is a function with parameters.
  • Inputs are supplied by data.
  • Parameters are stored by the model and changed by training.
  • Outputs are computed predictions.
  • A forward pass computes predictions using current parameters.
  • Prediction happens before loss, gradient, and update.
  • An architecture defines a family of possible functions.
  • Parameters choose one function from that family.

Notation

NotationMeaning
xinput
ytarget
y_hatprediction
thetamodel parameters
f(x; theta)model output for input x using parameters theta

Common mistakes

  • Treating input data as if it were learned by the model.
  • Forgetting that targets are supplied by the dataset.
  • Saying the model learns during the forward pass.
  • Talking about capacity without saying what function family the architecture can represent.

Before moving on

You should be able to compute y_hat = wx + b, name which values are data and which are parameters, and explain why training must compute predictions before it can improve them.