PERCEPTRON and PATTERN do not need normalized data. The perceptron is solved with a normal equation, not with gradient descent. Neural networks that use gradient descent or similar backprop algorithms would indeed require normalized, balanced data.