Train a binary perceptron from zero-initialized weights using ordered samples, step predictions, and error-correction updates.
