The fixed gate
A programmer chooses one threshold. Moving it changes behavior directly, but a one-feature gate cannot express a relation that depends on both A and B.
prediction = A >= threshold
Predict, train, explain
Rules are written. Boundaries are learned.
Six labeled signals enter the floor. First program a rigid gate. Then train a real perceptron and watch examples move its decision plane.
Prediction requiredWhich approach can represent a boundary that depends on both signals?
Predict the adaptable system
Inspect a system
Run the lesson
A programmer sets the rule. Examples shape the model.
Traditional programming executes an explicit condition. A learning algorithm still contains programmed instructions, but those instructions update parameters from errors. The examples do not write new if-statements; they change the boundary used on the next input.
The learned plane
The perceptron combines both signals. Its coefficients are not chosen sample by sample; labeled errors cause numerical updates.
prediction = sign(wA*A + wB*B + bias)
A worked error
For [0.7, 0.1] labeled 0, the initial activation is 0.13, so the model predicts 1. Error -1 reduces both weights and the bias.
delta wA = 0.25 * -1 * 0.7 = -0.175
The honest boundary
This narrow model learns one linearly separable task. Six correct samples do not imply broad human-like intelligence or guaranteed performance elsewhere.
training success != general intelligence
Misconception: AI is just more rules.
The update rule is programmed, but the final coefficients emerge from labeled mistakes. Inspect the epoch trace: each error changes the next decision.
Counterexample: move only A.
The gold point at low A and high B remains unreachable by a vertical threshold without breaking another case. The learned plane can tilt because it uses both features.
Interview answer: name the mechanism.
Traditional software maps inputs to outputs through explicit rules. Machine learning uses examples to estimate parameters, then applies the learned mapping to unseen inputs, with accuracy bounded by data and model assumptions.
Transfer the boundary.
After reaching 6/6, classify one unseen signal. Do not guess from its appearance; evaluate the learned activation.
unseen = [A 0.75, B 0.35]
activation = wA*A + wB*B + biasTrain to mastery, then apply the learned equation.