Did you know? A regression model can be equally accurate on average in two regions of a dataset while the uncertainty of individual observations differs dramatically. Heteroscedastic regression models this input-dependent variation instead of forcing every prediction to share one noise scale. It can predict a mean and a conditional variance, then fit both under a probability model such as Gaussian negative log-likelihood. But a predicted variance is not a licence to ignore difficult data, and a very low training loss does not guarantee trustworthy uncertainty.
The useful answer is to distinguish irreducible observation noise from uncertainty about the learned model, choose a probabilistic distribution appropriate to the outcome, train the mean and scale with numerically controlled loss, and validate predictions by both accuracy and conditional uncertainty coverage on independent data. When a model inflates variance to hide residual errors or drives it toward zero around memorised points, diagnose the objective and the data before declaring the uncertainty reliable.
This guide develops the mathematics, computes a full Gaussian example, explains reported optimisation pitfalls and transfers the reasoning to a new prediction problem. The worked measurements are invented teaching cases, not a claim that a real model was trained. The research links point to original published work and its limitations.
Contents
- Why one noise level is often the wrong question
- Aleatoric and epistemic uncertainty are not interchangeable
- Gaussian negative log-likelihood, line by line
- Worked numbers: why two variances give different NLLs
- Why a variance head can learn the wrong shortcut
- Parameterise positive variance without hiding units
- Accuracy and uncertainty need separate scorecards
- A checked two-region evaluation before declaring victory
- The beta-NLL response to a known optimisation problem
- When Gaussian variance is not the right uncertainty model
- Nine practical mistakes and their targeted repairs
- An implementation path from tiny examples to live evaluation
- Transfer exercise: sensor measurements with changed units
- Limitations and what a variance number cannot certify
- Frequently asked questions
- Reader pathways and primary research
Why one noise level is often the wrong question
Ordinary least-squares regression tries to learn a relationship between inputs and a numerical target. In a simple setting, y is treated as a predictable mean plus an error term whose variability may be modelled as constant. That constant-variance assumption is called homoscedasticity. It is sometimes adequate. But many real measurements become noisier as conditions change, and a single residual scale hides important differences.
Imagine estimating travel time from route length. Two short routes may vary by one minute, while two long routes can vary by fifteen minutes because of traffic and weather. A point prediction of twenty minutes and a point prediction of ninety minutes might both be close to observed averages, yet their uncertainty need not be comparable. Heteroscedasticity refers to variation in conditional noise across inputs, not merely different predicted means.
A heteroscedastic probabilistic regressor can learn two functions: a conditional expected value μ(x) and a conditional variance v(x) greater than zero. Under a Gaussian assumption, it models y conditional on x as Normal(μ(x), v(x)). The positive variance depends on the input and can respond to genuinely noisier regions. This is a model, not a claim that every real conditional distribution is Gaussian.
Why not simply attach a large warning band to every prediction? Because a uniform wide band can be unnecessarily vague for stable cases and dangerously narrow for difficult ones. A useful uncertainty model should explain where variation is high, where it is low and what evidence supports the difference. It must also distinguish this uncertainty from unfamiliar inputs where the model may not know what it is doing.
The original work by Kendall and Gal on aleatoric and epistemic uncertainty helped popularise learning input-dependent observation noise in deep visual prediction tasks. Later research points to specific training pitfalls. The distinction is useful for modern AI, but an estimated variance remains a testable prediction rather than an intrinsic truth attached to a sample.
Aleatoric and epistemic uncertainty are not interchangeable
Aleatoric uncertainty describes variability in observations or outcomes that remains even if the model parameters and task assumptions are well known. For example, two similar road trips can take different times because of circumstances not included in the input. Epistemic uncertainty describes uncertainty about the fitted relationship or parameters and can decrease as relevant information increases, at least under suitable assumptions.
Learning a variance head conditional on x commonly targets a form of aleatoric uncertainty. It does not automatically represent the model’s ignorance about regions with no training examples. A model can output a small variance for an out-of-distribution input because it has never learned how to recognise its lack of knowledge. Adding more expressive variance parameters can even make this failure easier to disguise.
Epistemic estimates may require different methods, such as suitable ensembles or approximate Bayesian approaches, evaluated against the application. The existing Monte Carlo dropout guide explains one way of generating model-dependent predictive variation and its limitations. It should not be treated as the same thing as a per-input Gaussian noise head.
Sometimes both uncertainties appear in the predictive distribution. A simple conceptual decomposition using the law of total variance is: total predictive variance equals expected conditional variance plus variance of the conditional mean across uncertain models. This decomposition depends on the chosen model and posterior approximation. It does not turn two arbitrary numeric scores into comparable probabilities.
A responsible report says which uncertainty is estimated, how it is validated and when its assumptions fail. Saying only that a model is 90% certain may be misleading if the number comes from a conditional noise variance measured in squared units rather than from a calibrated event probability.
Gaussian negative log-likelihood, line by line
Let μ be a predicted mean, v a strictly positive predicted variance and y the observed numeric target. A normal density has log probability involving the squared residual divided by variance and a normalising term involving log variance. Its per-observation negative log-likelihood is one half of [log(2π) + log(v) + (y−μ)^2/v].
The residual r=y−μ represents the point-prediction error in target units. Squaring r makes the numerator nonnegative. Dividing by v scales the penalty relative to the model’s claimed uncertainty; large errors are more plausible under a genuinely broad distribution. The log(v) term stops a model from making every variance arbitrarily large to reduce the residual term for all records.
Under a constant variance shared by every record, minimising the Gaussian negative log-likelihood with respect to mean parameters resembles least-squares training, because the residual-squared term is scaled by the same positive constant. Under input-dependent v(x), different records can receive different effective residual penalties. That flexibility is useful but creates the opportunity to underweight difficult training points.
The full log(2π) constant does not affect parameter gradients if nothing else changes, but it matters when reporting a true Gaussian negative log-likelihood comparable across implementations. Omitting the constant and labelling the remainder ‘exact Gaussian NLL’ causes a numerical mismatch. Likewise, predicting standard deviation σ and treating it as variance v changes the formula and its units.
Modern libraries can parameterise Gaussian distributions directly, but their input conventions vary. For a probabilistic implementation, a predicted variance must remain positive and numerically stable. Models often output a log variance or unconstrained value passed through a positive transformation. The engineering contract should record the chosen parameterisation and minimum safe scale.
Worked numbers: why two variances give different NLLs
Consider a fictional measurement y=12 and model mean μ=10 in the same units. The residual is r=2 and its square is 4. First let the predicted variance be v=1 squared unit. The Gaussian NLL is 0.5[log(2π)+log(1)+4/1]. Since 0.5log(2π)≈0.918939, the result is 0.918939+2 = 2.918939 nats.
Now suppose the same model predicts v=4 squared units. The residual penalty becomes 0.5×4/4=0.5 and the log variance contribution becomes 0.5log(4)≈0.693147. The total is 0.918939+0.693147+0.5≈2.112086 nats. A variance of four explains the large residual better under this Gaussian likelihood, but also pays a cost for being broad.
Try v=0.25 instead. The log contribution is 0.5log(0.25)≈−0.693147 while the residual penalty becomes 0.5×4/0.25=8. The NLL is 0.918939−0.693147+8≈8.225792 nats. A confidently narrow prediction that is two target units wrong receives a much larger loss.
Now change the residual. If y=μ=10 and v=1, NLL is 0.918939 nats. If v gets extremely small around a perfect training fit, the log variance can drive the continuous-density negative log-likelihood downward without bound. This is one reason variance collapse and numerical floors require attention. An infinitesimal predicted variance is not proof that future measurements will match exactly.
The NLL values are not probabilities of the observed point event; for continuous outcomes the probability of one exact numeric value is zero under a continuous density. NLL evaluates density at the observation, with conventions and target units affecting its absolute numeric level. Compare appropriately normalised outcomes rather than telling a reader that NLL 2.11 means a 2.11% error.
The example verifies two facts: a larger variance can be justified for a large residual, and extremely confident wrong predictions can be penalised sharply. It does not prove a neural model will discover the correct variance in deployment. Training dynamics and independent validation determine whether the predicted uncertainty is useful.
Why a variance head can learn the wrong shortcut
To see the optimisation incentive, let s=log(v). The per-observation Gaussian NLL, excluding the fixed normalising constant, is 0.5[s+r²exp(−s)]. Its derivative with respect to s is 0.5[1−r²/v], holding the mean prediction fixed. This expression tells us whether a local gradient step tends to increase or decrease log variance for a particular residual.
If r² is larger than v, the derivative is negative, so gradient descent increases s and therefore v. The model is encouraged to report more uncertainty for a surprisingly large residual. If r² is smaller than v, the derivative is positive, encouraging a smaller variance. For nonzero fixed residual the per-point optimum is v=r² in this simplified calculation.
The difficulty is that μ and v are often learned together from the same features. The mean head may make a large error in a region because its representation is poor. The variance head can respond by inflating predicted noise, which reduces the gradient pressure from that example on the mean. The model may settle into a poor mean with an apparently plausible large variance rather than learn a better conditional mean.
Seitzer, Tavakoli, Antic and Martius describe this concern in On the Pitfalls of Heteroscedastic Uncertainty Estimation with Probabilistic Neural Networks. They investigate circumstances where Gaussian NLL training yields poor but stable mean estimates and propose a beta-NLL alternative. Their experiments are evidence for particular models and datasets; the paper is not a licence to change losses without independent evaluation.
Another pathology is variance collapse around memorised training points. If a powerful model can fit particular observations almost perfectly and is not otherwise constrained, it can output tiny variances for them and achieve a highly favourable likelihood while being fragile away from training data. A numerical minimum variance, careful validation and appropriate regularisation can prevent pathological values, but the choice of floor must fit the target units and application.
Parameterise positive variance without hiding units
A neural network usually emits unconstrained real numbers. Variance cannot be negative, so the last transformation matters. Predicting s=log(v) and exponentiating can enforce positivity, but large positive s can overflow and very negative s can underflow. A softplus-based positive scale with a documented floor is another option. Neither transformation absolves the author from examining gradients and data scale.
Choose whether the model predicts variance v, standard deviation σ, log variance s, or log standard deviation. The formulas are not interchangeable. If s is log variance, v=exp(s), whereas if t is log standard deviation, v=exp(2t). An implementation that confuses these gives dramatically different uncertainty. A small unit test should verify the conversion using known numbers.
Standardising the target can help numerical conditioning, but you must transform uncertainty back correctly. If the original target is y=a+bz and a model predicts variance v_z for z, then the original target variance is b²v_z and its standard deviation is |b|√v_z. A mean prediction can look correct while the confidence interval is wrong by a factor of b if the variance rescaling is forgotten.
Similarly, predicted uncertainty needs units. Travel-time variance is expressed in minutes squared when the target uses minutes. Reporting ‘variance of four minutes’ confuses variance with standard deviation. A variance of four square minutes corresponds to a standard deviation of two minutes under the Gaussian assumption. Label charts and exported fields accordingly.
Logging and clipping should be explainable. A variance floor of 0.01 may be appropriate for a target measured in centimetres and absurd for a target measured in astronomical distances. Do not insert magic constants merely because an example notebook contained them. Quantify their effect on both training and out-of-sample calibration.
Accuracy and uncertainty need separate scorecards
Evaluate the mean prediction using metrics such as mean absolute error or root-mean-squared error, with consistent units and a representative independent test set. Evaluate probabilistic quality using proper scoring rules such as NLL, plus observed interval coverage and residual diagnostics appropriate to the task. A model that improves NLL by inflating variances may make point predictions worse, so report both.
Under a well-calibrated Gaussian predictive distribution, standardised residuals (y−μ)/σ should be compatible with a standard normal model in the population being evaluated. Plot or tabulate these values by predicted uncertainty level, input region and relevant subgroup. Heavy tails, skewness, changing bias or isolated outliers may show that the Gaussian family is inappropriate despite a smooth training curve.
Prediction interval coverage is particularly interpretable. If you construct nominal 90% intervals under the declared model and evaluate 200 independent comparable observations, the expected number inside under ideal calibration is around 180, though sampling variation matters. An observed 156/200=78% coverage suggests the model is too narrow or mis-specified for that set. It does not by itself identify which modelling assumption failed.
Check interval width at the same time. A model can achieve excessive coverage by issuing intervals too wide to help anyone. Evaluate the trade-off between width and coverage, and whether difficult subgroups receive appropriate uncertainty. A global 90% coverage rate can conceal 99% for one region and 75% for another.
Data shifts matter. A model trained on moderate traffic conditions may have no basis for representing rare disruptions. Its learned noise head cannot be assumed to extrapolate. The Gaussian process guide discusses a separate way to think about predictive distributions, kernels and uncertainty under model assumptions.
A checked two-region evaluation before declaring victory
Suppose a fictional weather-sensitive delivery estimator is evaluated on 100 independent journeys: 50 in a stable region and 50 in a variable region. For each journey it issues a nominal 90% Gaussian prediction interval. In the stable region, 47 of 50 observed durations fall inside. In the variable region, only 35 of 50 fall inside.
The observed coverage in the stable region is 47/50=94%. In the variable region it is 35/50=70%. Across all 100 journeys, 82 fall inside, so global coverage is 82/100=82%. Reporting only ‘82% coverage’ hides the severe variable-region failure, and reporting the stable result alone would make the model seem overly reliable.
Imagine a rival model produces 44 of 50 covered in the stable region and 44 of 50 in the variable region, or 88% in each and 88% overall. That is closer to the nominal 90% in this sample. But if its intervals are five times wider, the improved coverage may be operationally useless. The evaluation must include median or mean interval width in original target units and the consequences of incorrect intervals.
The denominators matter. If one subgroup had only five observations, a shift of one case would move coverage by twenty percentage points. Do not present small-sample percentages as precise population properties. Use uncertainty intervals, suitable resampling or more data to make stronger claims, especially for safety-critical use.
This fictional two-region exercise illustrates an evaluation procedure rather than a result for a real delivery service. Its lesson is to connect probabilistic predictions to events that can be checked after the fact, rather than judging uncertainty solely by how reasonable the printed sigma value looks.
The beta-NLL response to a known optimisation problem
The 2022 pitfalls paper motivates a modification called beta-NLL, in which the per-sample Gaussian NLL contribution is reweighted using the model’s predicted variance raised to a beta exponent. The authors investigate how this changes the optimisation pressure that otherwise can allow high-variance examples to exert too little influence on mean learning.
The practical interpretation is subtle. It is not simply another way to report a calibrated probability distribution. The training objective may no longer be the exact ordinary Gaussian negative log-likelihood. Its benefits must be assessed using held-out point accuracy and probabilistic quality, and the effect of the weighting convention must be stated precisely.
A practitioner should begin with the original NLL baseline and reproduce a small, well-defined pathological case before introducing beta-NLL. Check that the implementation’s weighting behaves as intended and that any detached factors, reductions or gradient conventions match the selected reference implementation or derivation. Do not silently call a changed objective ‘maximum likelihood’.
The paper’s empirical findings are useful because they separate bad optimisation behaviour from the simple mathematical attractiveness of likelihood. A theoretically familiar loss can still be hard to optimise jointly for mean and variance with shared neural parameters. A stable training curve is not proof that both parameters are well learned.
More complexity is not always the repair. Better target scaling, more relevant features, a different conditional distribution, variance constraints or better train-validation splits may address the actual issue more directly. Compare these options using a held-out dataset and a decision-specific metric before standardising one loss across every project.
When Gaussian variance is not the right uncertainty model
A Gaussian conditional distribution is symmetric and defined by mean and variance. Many real outcomes are nonnegative, skewed, multimodal, bounded or heavy-tailed. Travel times can have a long right tail; counts are discrete; proportions live between zero and one. A symmetric Gaussian prediction interval can assign impossible negative values or fail to reflect multiple plausible outcomes.
For counts, an appropriate Poisson or negative-binomial model may be more natural, with important distinctions in how mean and variance relate. For positive continuous values, log-normal or gamma-like models can be useful when the assumptions match. For strongly heavy-tailed errors, a Student-t likelihood may be less sensitive to extreme observations. Each distribution comes with its own parameters and checks.
Quantile regression avoids committing to a full parametric conditional density when the task needs particular conditional quantiles. The quantile regression guide explains pinball loss and asymmetric decisions. It is not the same as a Gaussian model with a changing variance, but it can provide useful ranges when normality is implausible.
Distribution-free predictive coverage procedures such as conformal prediction address a different question: coverage guarantees under specified exchangeability assumptions, with caveats under distribution shift. The choice between a learned variance head, calibrated quantiles and conformal intervals depends on the output and evidence required. These methods can also be combined thoughtfully, but combination does not remove their assumptions.
An engineer should begin by plotting real residuals against inputs and fitted means, checking the target’s support and deciding which consequences matter. A model chosen because its loss function is convenient may describe the wrong world even when its gradient calculations are correct.
Nine practical mistakes and their targeted repairs
First, predicting standard deviation but passing it where a library expects variance squares or halves the uncertainty by mistake. Repair this with a small hand-calculated NLL test and explicit units. Second, allowing a variance of zero or negative values, which makes log or division undefined. Repair by choosing a stable positive parameterisation and a justified floor.
Third, reporting uncertainty in scaled target units after transforming the mean back to the original scale. Repair by applying the squared scale factor to variance. Fourth, treating aleatoric variance as epistemic ignorance. Repair by evaluating extrapolation and unfamiliar inputs separately, including suitable model-uncertainty methods when the application needs them.
Fifth, training the variance head to absorb errors caused by a weak mean head. Compare point accuracy, residuals and variance patterns, not likelihood alone. Sixth, choosing a variance floor by intuition without considering target units or a change in measurement equipment. Document scale and conduct sensitivity checks.
Seventh, using random record splits when several observations come from the same person, device or site. Leakage can make both mean accuracy and uncertainty calibration look better than they will be in deployment. Use independent splits matched to the future prediction job. Eighth, averaging subgroup coverage into one pleasing number. Report regions and populations relevant to error cost.
Ninth, turning a favourable training NLL into a statement that the model is safe for external decisions. Probability density fit is not a guarantee of correct action or trustworthy causal reasoning. Repair by defining the use case, checking harms and placing independent guardrails around consequential predictions.
The common theme is that a variance output is only one part of a prediction contract. The owner needs correct mathematics, honest data boundaries, stable optimisation, relevant calibration and a clear policy for what the user should do when the model says uncertainty is high.
An implementation path from tiny examples to live evaluation
Start with a synthetic dataset whose true noise process is known to the experimenter. For example, let the conditional mean follow a simple line while residual spread increases with x. Fit a constant-variance baseline and an input-dependent variance model. Compare whether the second captures genuinely different uncertainty while preserving an accurate mean. Synthetic truth helps debug an estimator but does not demonstrate performance on real data.
Then unit-test the Gaussian loss with one observation, known residual and several variance values. Verify that the loss for residual two and variance four is approximately 2.112086 nats under the full formula. Check gradients numerically for a small range of log variance values. Verify that clipping or flooring does not silently mask invalid values.
Next train on a genuine dataset with feature preprocessing fitted only on the training partition. Separate model selection from final testing. Keep the same data split, backbone and optimiser when comparing uncertainty objectives, unless a deliberate factorial design changes more than one variable. Report the chosen parameterisation, minimum variance, batch reduction and random seeds.
Inspect residuals versus the input, predicted scale and actual absolute error. A conditional noise model should generally assign more uncertainty where outcomes are intrinsically variable, but do not expect perfect per-instance correspondence: realised error is one draw from a distribution. Evaluate aggregate conditional behaviour over comparable observations.
Finally define a bounded application rule. A delivery scheduler might offer a range and suggest buffers for volatile routes. It should not label every uncertain route ‘unsafe’ unless the operator has specified a real decision threshold. The statistical estimate supports decisions but does not substitute for the user’s domain knowledge.
Transfer exercise: sensor measurements with changed units
Change the application to a sensor predicting room temperature. The model was trained on Celsius outcomes and predicts mean 22°C with variance 4°C², so the predicted standard deviation is 2°C. Assume for this teaching exercise that its Gaussian approximation is otherwise adequate and that the target transformation is purely linear.
An engineer exports values in Fahrenheit using F=32+(9/5)C. The mean becomes 32+(9/5)×22=71.6°F. Variance must scale by (9/5)²=3.24, so the Fahrenheit variance is 4×3.24=12.96°F². The standard deviation becomes √12.96=3.6°F. Simply keeping variance four after the unit conversion would produce a wrong uncertainty report.
Now imagine the sensor’s readings grow noisier when the device heats above a threshold. A constant-variance model would conceal that conditional change. A heteroscedastic model might learn larger variance for the hotter input region if the training set contains representative observations. But if the sensor is operated at a temperature never observed during training, predicted variance may not capture the new failure mode.
Add a numerical evaluation: the model issues nominal 90% intervals for 40 test readings in the hotter region, and only 29 true values fall inside. The observed coverage is 29/40=72.5%, well below the nominal label in this small set. This warns about the uncertainty model or assumptions, though it does not identify whether the cause is calibration, distribution shift, sensor faults or limited sample size.
The transfer solution combines unit-correct variance transformation, conditional evaluation and an engineering escalation when the use case changes. It does not mistake a precise sigma number for proof of reliable monitoring.
Limitations and what a variance number cannot certify
Heteroscedastic regression is a flexible probabilistic model, not a universal explanation of uncertainty. Gaussian likelihood assumes a form for the conditional distribution that can fail under skewness, discrete outcomes, multimodality or heavy tails. A more complicated network does not automatically fix a wrong distribution family.
A correctly calculated NLL can still reward poor mean predictions under difficult optimisation conditions. Original research documents this hazard, so practitioners should compare point prediction and uncertainty quality. A variance head should not be trusted merely because the training process converged without numerical errors.
Conditional variance from a neural network is not an independent estimate of ignorance outside the training distribution. New environments and rare events can break the learned relationship. Distinguish the output’s numerical precision from evidential confidence in the model itself.
Some decisions need well-calibrated event probabilities, upper bounds on rare losses or guarantees under specified assumptions, not Gaussian variance alone. Define the required property first. When no appropriate evidence supports such a property, the system should abstain or ask for human review rather than print a comforting interval.
All worked values in this article are invented for explanation and checked under their stated assumptions. Nothing here claims that a real sensor, delivery service or neural network was trained or validated during preparation of this guide.
Frequently asked questions
What is heteroscedastic regression? It models the variability of a target as changing with input conditions, often by predicting both a conditional mean and a conditional variance.
How does it differ from ordinary least squares? A common least-squares model uses an unweighted squared residual objective and can be paired with a constant-variance assumption. Heteroscedastic probabilistic modelling lets variance depend on input and changes the likelihood weighting.
Why must predicted variance be positive? The Gaussian density includes log variance and division by variance; zero or negative values are invalid. Use a stable positive parameterisation.
Does a larger variance mean the mean prediction is worse? Not necessarily. It can indicate genuinely noisy outcomes even when the expected value is accurate. But models can also inflate variance to hide mean errors, so evaluate both.
What is variance collapse? It is a pathological tendency toward extremely small estimated variance, often around closely fit training observations, creating unstable or misleading likelihood behaviour.
Is Gaussian NLL the same as RMSE? No. Gaussian NLL evaluates a predictive distribution under its assumptions, including a variance term. RMSE describes point-prediction squared error on a dataset.
Can an estimated sigma detect unseen situations? Not reliably by itself. Input-dependent aleatoric uncertainty and model ignorance are distinct; evaluate unfamiliar inputs with appropriate methods.
Does beta-NLL always improve performance? No. It is a researched alternative motivated by identified optimisation pitfalls, not a universal guarantee. Test it against a reliable baseline.
How should predictions be presented to users? With clear target units, honest uncertainty interpretation, validation context, limitations and an action policy that matches the task.
Reader pathways and primary research
If you are learning machine learning, derive Gaussian NLL, compute it for residual two under variances one, four and one quarter, then check the direction of the log-variance gradient. The arithmetic reveals the core incentives more clearly than a diagram of a black box.
If you build regressors, compare constant and conditional noise models on truly held-out examples. Track mean error, probabilistic score, subgroup coverage and interval width. Use unit tests for positive variance, target standardisation and exported scale. Read Gaussian process regression and Monte Carlo dropout for different uncertainty mechanisms.
If you audit risk-sensitive predictions, ask what observations support each uncertainty claim, how changes in environment are handled and whether the output describes target noise or model ignorance. Do not accept a confidence band detached from calibration evidence or units.
Primary references: Kendall and Gal (2017), What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?; Seitzer, Tavakoli, Antic and Martius (2022), On the Pitfalls of Heteroscedastic Uncertainty Estimation with Probabilistic Neural Networks; and Gurevich and Stuke (2017), Pairing an Arbitrary Regressor with an Artificial Neural Network Estimating Aleatoric Uncertainty. These are source studies, not claims of universal effectiveness.
For a broader foundation, visit How AI Works and the education learning-transfer guide.