Data is linearly separable when a linear decision boundary can classify its classes without error. This assumption enables the perceptron to converge after finitely many iterations. section 5.2
The perceptron treatment connects separability to implementation details:
- The algorithm starts with zero parameters and updates them whenever a training example is misclassified. section 5.2
- When the training examples are linearly separable, these updates converge after a finite number of iterations. section 5.2
- Input order affects the learned parameters, so shuffling the training data can be useful. section 5.2
Perceptron classification
A perceptron implemented from scratch trains on iris feature data and predicts test labels from the sign of its linear decision boundary. section 5.2
