#learning-rule

Run the perceptron algorithm (step size 1, initial weights $\mathbf{w} = (0, 0, 0)$) on the augmented points $(1, x_1, x_2)$, cycling through them in this order: negative $(-1, 0)$, negative $(2, -2)$, positive $(1, 0)$. A point counts as misclassified when $y\,(\mathbf{w} \cdot \mathbf{x}) \le 0$, and then $\mathbf{w} \leftarrow \mathbf{w} + y\,\mathbf{x}$. How many updates happen before a full pass makes no mistake?