Perceptrons as Simple Learners
A one-layer learner gives a simple yes-or-no delivery warning.
Read the full lesson and open its transcript below. A narrated video will be added after production.
What you will learn
A one-layer learner gives a simple yes-or-no delivery warning.
A practical question
A small perceptron takes predeparture features and produces a yes-or-no warning. It multiplies inputs by weights, combines them with a bias, and applies a decision rule.
The useful idea
During training, a mistaken positive example can push the weights toward producing a positive response for that input. A mistaken negative can push them away. Trace one update with a tiny numeric example before running thousands of trips. This learner draws a linear boundary in its feature representation, so it cannot express every route-weather interaction by itself.
Work through the example
Compare its warnings with a simple linear classifier and inspect the same held-out deliveries. Scaling and feature availability still matter. A neural name does not exempt the method from leakage checks. A perceptron can be understood one mistaken prediction at a time. Start with a small route record and its known label. If the current weights predict the wrong class, adjust them in the direction that would have favored the correct answer. Trace the changed score on that example and one other trip. An update that helps one case may hurt another, so repeated passes and independent evaluation still matter. This exercise reveals the learning mechanism without pretending the final weights explain traffic causally. Now check the idea on a concrete example: Trace one update after a mislabeled training prediction. A possible trap is worth naming: A single linear decision surface cannot represent every nonlinear pattern. The answer follows from that evidence: A perceptron combines weighted inputs and updates weights on mistakes under its training rule. After one perceptron update, score both the corrected trip and a previously correct trip. The second score shows why learning is not just fixing mistakes one at a time. Save the feature values so the sign and size of the update can be explained.
Limits and next step
Next, we'll replace the simple mistake rule with a differentiable loss and see how a gradient guides an update.
Worked example
Trace one update after a mislabeled training prediction.
Quick check: test your understanding
Try it yourself: Consider this case: Trace one update after a mislabeled training prediction. Explain how you would check the result. In your answer, address this limitation: A single linear decision surface cannot represent every nonlinear pattern.
Check the reasoning
A perceptron combines weighted inputs and updates weights on mistakes under its training rule. A single linear decision surface cannot represent every nonlinear pattern.
What is the main idea behind perceptrons as simple learners?
A perceptron combines weighted inputs and updates weights on mistakes under its training rule.
How is it used in the example?
Trace one update after a mislabeled training prediction.
What limitation should be checked?
A single linear decision surface cannot represent every nonlinear pattern.
Further reading
Machine Learning — Tom M. Mitchell.
Original ToolsInVision teaching examples. Editorial approach · Suggest a correction
Accessible referenceLesson 53 transcript312 words · Narration ready for the upcoming videoOpen transcript
A small perceptron takes predeparture features and produces a yes-or-no warning. It multiplies inputs by weights, combines them with a bias, and applies a decision rule.
During training, a mistaken positive example can push the weights toward producing a positive response for that input. A mistaken negative can push them away. Trace one update with a tiny numeric example before running thousands of trips. This learner draws a linear boundary in its feature representation, so it cannot express every route-weather interaction by itself.
Compare its warnings with a simple linear classifier and inspect the same held-out deliveries. Scaling and feature availability still matter. A neural name does not exempt the method from leakage checks. A perceptron can be understood one mistaken prediction at a time. Start with a small route record and its known label. If the current weights predict the wrong class, adjust them in the direction that would have favored the correct answer. Trace the changed score on that example and one other trip. An update that helps one case may hurt another, so repeated passes and independent evaluation still matter. This exercise reveals the learning mechanism without pretending the final weights explain traffic causally. Now check the idea on a concrete example: Trace one update after a mislabeled training prediction. A possible trap is worth naming: A single linear decision surface cannot represent every nonlinear pattern. The answer follows from that evidence: A perceptron combines weighted inputs and updates weights on mistakes under its training rule. After one perceptron update, score both the corrected trip and a previously correct trip. The second score shows why learning is not just fixing mistakes one at a time. Save the feature values so the sign and size of the update can be explained.
Next, we'll replace the simple mistake rule with a differentiable loss and see how a gradient guides an update.