Summary and Revision Notes

Key ideas

  • A linear score is a weighted sum plus a bias.
  • Weights control feature contribution.
  • Bias shifts the score.
  • Raw scores are not automatically probabilities.
  • A score vector contains one score per class.
  • A binary linear decision boundary is where x . w + b = 0.
  • In two dimensions, that boundary is a line.
  • Linear models are useful but limited by linear boundaries.

Notation

NotationMeaning
xinput feature vector
wweight vector for one score
bbias
x . w + bone linear score
xW + bvector of scores

Common mistakes

  • Treating raw scores as probabilities.
  • Forgetting the bias term.
  • Confusing number of input features with number of classes.
  • Assuming a linear model can represent every pattern if trained long enough.

Before moving on

You should be able to compute x . w + b, identify weights and bias, predict score shapes, and explain why nonlinear activations are needed after linear models.