Why Neural Nets Fit Noise
A network that can fit unseen noise isn't memorizing exceptions — it's because high-capacity models can interpolate the training data exactly, which forces the learned function to be smooth enough to generalize on real patterns.
A complete interactive classroom, not just a preview.
Start when you are ready to enter this Stage's 12 scenes and explore, respond, and learn as you go.
Why can a network memorize random noise it has never been trained on, and what does that say about generalization?
- interpolation-vs-approximation
- Distinguish models that approximate training data from models that pass exactly through every training point.
- overfitting-misconception
- Overfitting is not merely 'fitting noise' but failing to generalize; fitting noise is sometimes compatible with generalization.
- capacity-and-overparameterization
- Modern neural networks have enough parameters to interpolate any labeling of the training set, including random labels.
- implicit-bias-smoothness
- Gradient descent on neural networks implicitly prefers low-complexity, smooth solutions among all functions that fit the data.
- noise-as-stress-test
- The ability to fit random labels is a probe for capacity; it doesn't mean real labels are memorized the same way.
If a model can fit random noise, it must be memorizing everything and generalizing badly.
Show that capacity to interpolate is separate from generalization; a model can fit noise on one dataset and still generalize on another because optimization biases solutions toward simplicity.
Overfitting means the model learned the training set too well.
Clarify that overfitting is a gap between train and test performance, not the absolute ability to fit train data.
More parameters always mean more overfitting.
Explain that overparameterized networks trained with gradient descent often generalize well because the optimizer can't reach arbitrary solutions — it finds simple ones.
- Basic idea of supervised learning and train/test split
- Concept of a loss function being minimized
- Familiarity with the term 'overfitting'
- Specific architectural proofs (NTK details)
- Formal statistical learning theory bounds
- Adversarial robustness
- Hyperparameter tuning recipes
- Learner explains why a network that fits random labels can still generalize on real data.
- Learner distinguishes interpolation from approximation in their own words.
- Learner identifies implicit bias toward smoothness as the mechanism connecting capacity to generalization.
- Apply this framing to judge whether surprising training-set results (perfect train accuracy, memorization probes) actually predict test-set behavior.
Learners familiar with basic supervised learning, training vs. test sets, and the idea that models learn from data. No deep optimization theory required.
- 01The PuzzleslideOrientationObserve
Pose the core question: networks fit random labels perfectly, yet generalize on real data. Why?
- Zhang et al. showed CNNs can drive training error to zero on random labels
- Yet the same architectures generalize on real images
- The contradiction hints that training error is not the right measure of learning
- 02Interpolation vs. ApproximationinteractiveModel buildingConstruct
Let learners drag a flexible curve through scattered points and see that exact fit is always possible when the model is flexible enough.
- An interpolating curve passes through every point
- An approximating curve only needs to be close
- With enough flexibility, every labeling becomes reachable
- 03What Overfitting Really MeansslideMisconception repairExplain
Reframe overfitting as a generalization gap, not as 'fitting the training set'.
- Overfitting = train accuracy high AND test accuracy low
- Zero training error is necessary but not sufficient for overfitting
- Two networks can both fit the data perfectly and behave very differently on test data
- 04Check Your IntuitionquizAssessmentChoose
Commit to an interpretation of two identical training-accuracy curves with different test behavior.
- Identify which framing predicts test performance
- Spot the overfitting red herring
- 05Capacity Without ConstraintslideModel buildingObserve
Show how overparameterized networks can realize any labeling of the training set in principle.
- Parameter count often exceeds training sample count
- Universal approximation: enough width reaches any continuous function
- Random labels are just another target function to interpolate
- 06The Hidden Constraint: Gradient DescentslidePredictionPredict
Predict what kinds of solutions an optimizer can actually find among all functions that fit.
- Many functions interpolate the data
- Gradient descent does not visit them uniformly
- Ask: which one will training actually pick?
- 07Simulating Implicit BiasinteractiveModel buildingObserve
Visualize gradient descent picking a smooth path through points even when jagged paths also fit.
- Watch the chosen curve evolve over training steps
- Compare final curve to alternative interpolants
- See the optimizer prefer low-frequency, smooth solutions
- 08Smoothness as a PriorslideMisconception repairExplain
Connect implicit bias to generalization: smooth functions align with the structure of natural data.
- Natural images and language have low-frequency structure
- Smooth-biasing solutions exploit that structure on unseen points
- Random labels lack that structure — smoothness stops helping there
- 09Noise as a Probe, Not a VerdictslideSynthesisExplain
Reframe the original puzzle: fitting noise is a capacity test, not evidence of bad generalization.
- Random-label experiments measure capacity, not memorization habits
- Real data has structure that smoothness exploits
- The two settings ask different questions of the model
- 10Apply the FrameslideApplicationApply
Apply the capacity-vs-bias lens to a new scenario: a model with perfect train accuracy on a small dataset.
- Decide whether train accuracy tells you about test performance
- Look for evidence of structure the model could exploit
- Diagnose overfitting by train-test gap, not train error alone
- 11Final CheckquizAssessmentExplain
Verify learners can explain why interpolation on noise is compatible with generalization elsewhere.
- Distinguish capacity from generalization
- Identify the role of optimization bias
- 12Key TakeawaysslideSynthesisExplain
Summarize the three-part answer: capacity, optimization bias, structure of real data.
- Overparameterized networks can interpolate any labeling
- Gradient descent biases solutions toward smooth functions
- Generalization comes from matching that bias to the structure of real data
Discussion threads for a Stage aren't available yet.
