Bias, Variance, and Learning-Curve Diagnostics

An average validation score can hide very different failure mechanisms. The useful question is not only which model scored highest, but why each candidate failed.

Each of the four patterns below has been measured in published work, on datasets the field already knows by name. The measurement is what separates a diagnosis from a label, and every one of these numbers is checkable.

The classical bias–variance vocabulary is most useful when each term is tied to observable evidence rather than used as a label for any disappointing result. Four mechanisms are worth keeping apart.

Approximation or specification limits: the chosen features, objective, or model family cannot represent the relevant relationship well enough, even on training data. Estimation sensitivity: the fitted system changes materially when the sample, random seed, initialization, or preprocessing choices change. Observation and label noise: measurement ambiguity, delayed outcomes, inconsistent labels, or unobserved causes place a floor beneath achievable error. Deployment mismatch: the evaluation population, workflow, or time period differs from the environment where the decision will be used.

The third of those is usually asserted and left there. It has been measured. Ten of the most-used vision, language and audio benchmarks have wrong labels in their test sets, and someone counted them. Northcutt and colleagues reported the count at NeurIPS in 2021. The abstract: “Errors in test sets are numerous and widespread: we estimate an average of at least 3.3% errors across the 10 datasets, where for example label errors comprise at least 6% of the ImageNet validation set.”

The consequence is not a caveat about precision. It reverses a capacity decision. On corrected ImageNet labels, ResNet-18 overtakes ResNet-50 once the prevalence of originally mislabelled test examples rises by 6%. The smaller model wins because the reference was wrong, not because capacity was wrong. That is exactly the confusion this map exists to prevent. And the floor is itself contestable in the right way. A Google team hand-reviewed every remaining error of top ImageNet models and found that nearly half were not errors at all.

Training and validation behavior provide clues, not final diagnoses. Controlled changes are needed before assigning a cause.

Bias-like pattern. Both training and validation performance remain weak, and extra data produces little improvement. Check whether the target and available features make the task learnable, compare against a richer representation or model family, and inspect residual structure rather than only the aggregate metric. A linear model facing a strong nonlinear interaction is the standard shape of this failure.

Variance-like pattern. Training performance is strong, but fresh-sample performance is weaker or unstable across reasonable resamples. Measure variation across grouped splits, seeds, and relevant environments; test regularization, capacity reduction, ensembling, or better data coverage; and audit leakage before interpreting a surprisingly large train–validation gap. A deep tree memorizing rare identifiers is the standard shape.

Mismatch pattern. Training and ordinary validation look strong, and then the same model meets a new sample drawn the same way. In 2019 Recht and colleagues rebuilt the CIFAR-10 and ImageNet test sets, following the original collection procedures, and ran the field's own models against them. The abstract: “We evaluate a broad range of models and find accuracy drops of 3% - 15% on CIFAR-10 and 11% - 14% on ImageNet.”

What happened next is the lesson inside the lesson. An independent MIT group re-analysed the ImageNet-v2 drop in 2020. After correcting for statistical bias in the replication procedure, only about 3.6% ± 1.5% of the original 11.7% ± 1.0% drop remained unaccounted for. Most of the gap was an artefact of how the replication was built. A residual survived. A gap is a question, and the controlled follow-up is what answers it. So rebuild the split to match the intended transfer question, inspect changing prevalence, measurement, workflow, and label availability, and do not describe domain shift as variance without supporting evidence.

A data learning curve plots performance against the amount of representative training evidence. A capacity curve varies model flexibility, regularization, or another complexity control while preserving the evaluation protocol.

The shape of the first has been measured rather than assumed. In 2017 Hestness and colleagues at Baidu Research tracked generalization error against training-set size in machine translation, language modelling, image processing and speech recognition. They found power-law scaling in all four. Architecture work moved the line down without tilting it. The abstract: “Further, model improvements only shift the error but do not appear to affect the power-law exponent.” The curve moves down, not sideways.

DeepMind's Hoffmann and colleagues reached a compatible conclusion in 2022, and then spent real compute on it. Over 400 language models, from 70 million to over 16 billion parameters, trained on 5 to 500 billion tokens. Then one controlled test of the fitted relationship: Chinchilla, 70B parameters trained on 4× more data, against Gopher, 280B parameters, at equal compute. That is what reading a curve and then deciding what to buy looks like when it is run as an experiment instead of an intuition.

But a curve is evidence to read, not a line to extend. More data can make a model worse. Nakkiran and colleagues found the regime where it does, in a paper titled Deep double descent: where bigger models and more data hurt. The abstract: “Furthermore, our notion of model complexity allows us to identify certain regimes where increasing (even quadrupling) the number of train samples actually hurts test performance.” More data made the model worse, in a measured regime, on purpose. And the counterexample is not a curiosity of deep networks. An independent TU Delft group proved at NeurIPS in 2019 that standard empirical-risk minimizers can behave non-monotonically at any training-set size, across classification, regression and density estimation.

So the reading rule is local. If training and validation performance converge at an unsatisfactory level, more of the same data may offer little benefit. If the validation curve keeps improving and the gap narrows, collecting more representative data is still plausibly worth paying for. That is a claim about this curve, to be checked at the next increment, not a law. Curves should include uncertainty across appropriate resamples or runs, because one smooth line can conceal instability, subgroup failures, or a split that leaks information.

Change one source of capacity or evidence at a time, then read the entire pattern rather than selecting the most flattering point.

1. Freeze the evaluation contract. Keep units, split logic, metric definitions, thresholds, and test access fixed while diagnosing the training process.

2. Establish floor and ceiling references. Include a simple baseline, an oracle-like diagnostic where possible, and estimates of label or measurement reliability. The field's most-quoted ceiling shows what building one actually involves. ImageNet's own challenge paper had an annotator, A1, label a 1,500-image sample of the ILSVRC2012–2014 test set. A1 reached 5.1% top-5 error, against GoogLeNet's 6.8% on the same sample. A second, less-trained annotator, A2, reached only about 12.0%. Same task, and the ceiling moved with how much the human had trained — A1 on 500 images, A2 on 100. A human ceiling is a measurement with a protocol, not a constant. It was later rebuilt. Shankar and colleagues ran a year-long experiment with five trained human labelers on 40,000 ImageNet and ImageNetV2 images, published at ICML in 2020. Their abstract: “Only the latest models from 2020 are on par with our best human labeler, and human accuracy on the 590 object classes is still 4% and 10% higher than the best model on ImageNet and ImageNetV2, respectively.”

3. Vary evidence and capacity separately. Construct learning curves over data volume and controlled sweeps over model flexibility or regularization, one at a time.

4. Repeat under relevant perturbations. Use multiple seeds, grouped resamples, time windows, or domains that match plausible deployment variation. Do not mistake the spread across cross-validation folds for that variation. Bengio and Grandvalet proved in 2004 that no universal unbiased estimator of the variance of K-fold cross-validation exists. Varoquaux measured what that costs in practice in 2018: at a sample size of 100 the true error bar is about ±10%. The abstract of that paper states the consequence flatly: “The standard error across folds strongly underestimates them.”

5. Select the next falsifiable experiment. Choose the cheapest test that separates competing explanations for the observed gap or plateau.

A bias-like pattern may justify richer features, a different objective, or a more expressive model. A variance-like pattern may justify better coverage, regularization, ensembling, or a simpler decision rule.

If curves disagree across sites or time periods, work on evaluation design and shift analysis before generic capacity tuning. The ImageNet-v2 sequence is what that work looks like when someone does it: a measured drop of 11% - 14%, then a re-analysis leaving about 3.6% ± 1.5% unexplained. If labels are unreliable, improving the reference process may outperform another model search. A test set with at least 6% of its labels wrong can rank a smaller model above a larger one, for reasons that have nothing to do with capacity.

Record the diagnosis, the competing explanations, and the expected signature of the next experiment. That turns improving the model into a sequence of testable decisions rather than an unstructured hyperparameter hunt.