How Examples Teach Computers
A computer learns from examples by turning each mismatch between its prediction and the correct answer into a small parameter adjustment; after many adjustments, it captures a pattern that transfers to new cases.
A complete interactive classroom, not just a preview.
Start when you are ready to enter this Stage's 4 scenes and explore, respond, and learn as you go.
How does showing a computer examples improve its ability to handle new cases?
A photo app tags your friends after seeing just a few tagged photos.
If it simply memorized those tagged photos, it would fail on new angles, lighting, and expressions; so something more than storage must be happening.
An interactive simulator where moving a model's parameters lowers its error on examples, making visible that learning is tuning, not copying.
A computer learns from examples by turning each mismatch into a small correction that tunes a flexible pattern, which then transfers to new situations.
- deep neural network architecture
- data collection pipelines
- model deployment
- algorithmic fairness
- 01The Name That Photo TrickslideSlot 1Hook
Start with a familiar surprise: after seeing only a handful of tagged photos, the app begins labeling brand-new pictures correctly.
- Only a few examples are used
- The photos are new, not exact copies
- The computer must be doing something more than memorizing
PhenomenonA photo app labels new pictures of a person after seeing only a handful of tagged examples.
QuestionWhat could the computer be doing with those examples to recognize a face it has never perfectly seen?
- 02Is It Just a Giant Memory?slideSlot 2Tension
Consider the tempting idea that the computer stores every example and matches new inputs against them.
- Computers are good at storing data
- Exact matching would fail with new lighting and angles
- Something deeper must create the ability to generalize
PredictionPerhaps the computer compares each new input to stored examples and returns the closest match.
Tempting intuitionSince computers store data so well, learning seems like saving examples and then doing a lookup — no real pattern-making needed.
- 03Error-Adjust SimulatorinteractiveSlot 3Reveal
Move the slope and intercept knobs to fit a set of example points and watch the total error change. The knobs are the computer's parameters.
- Each example gives a prediction and an error
- Turning the knobs changes the pattern
- Small error-driven corrections are what learning looks like
EvidenceWhen the line misses the example points, the error number is high; moving the knobs in the right direction makes the error shrink on the examples.
ConclusionLearning from examples is not copying them; it is repeatedly turning mistakes into parameter corrections.
Mechanism- 1Each example produces a prediction and an error: predicted answer minus actual answer.
- 2The computer nudges its parameters in the direction that reduces that error.
- 3Repeating many small nudges across examples shapes a pattern that captures the underlying relationship.
- 04The Real Sign of LearningslideSlot 4Takeaway
Test the idea in a nearby situation: after learning from a few tagged photos, a spam filter can sort new emails it has never seen before.
- New data is the real test
- Small corrections create transferable patterns
- Error on unseen examples matters more than memory
TransferAfter seeing tagged photos, the same error-adjustment rule explains how a spam filter learns to sort messages it has never met.
Expected inferenceIf a model only memorized training examples, it would fail on new inputs; success on unseen input is the real sign it has learned a pattern.
Discussion threads for a Stage aren't available yet.
