---
title: Expected Error Reduction in Active Learning
url: https://www.emergentmind.com/topics/expected-error-reduction-eer
type: topic
---

# Expected Error Reduction in Active Learning

Searching arXiv for the cited papers and related usage of “Expected Error Reduction (EER)” to ground the article.
Expected Error Reduction (EER) is an active learning criterion in which the next query is selected to maximally reduce the model’s expected future error on an unlabeled or validation set. In the formulation revisited for deep learning, the central principle is to choose the candidate whose label, once acquired, is expected to make the model best on the future evaluation distribution [2211.09283]. The acronym is, however, overloaded across adjacent literatures: in biometrics it commonly denotes equal error rate, and in factorial experiments it denotes experimentwise error rate [1906.06262] [2104.00732] [2309.12237] [2507.13621]. That ambiguity matters because the active-learning notion of EER is a query-selection objective, whereas the biometric and statistical notions are evaluation or error-control metrics.

## 1. Core active-learning definition

In active learning, the classic Expected Error Reduction idea from Roy and McCallum is: pick the unlabeled candidate whose label, if acquired, would maximally reduce the expected future error on a validation or unlabeled set [2211.09283]. In the paper’s notation, the candidate sample is \(x_i\) with random label \(Y_i\), and for a validation point \(x_j\) with label \(Y_j\), the general EER score is

\[
\text{Score}_i =
\sum_{j=1}^{n_\text{val}}
\min_{a \in \mathcal{A}}
\mathbb{E}_{y_j \sim Y_j \mid \mathcal{D}}
\big[\ell(y_j,a)\big]
-
\mathbb{E}_{y_i \sim Y_i \mid \mathcal{D}}
\left[
\sum_{j=1}^{n_\text{val}}
\min_{a \in \mathcal{A}}
\mathbb{E}_{y_j \sim Y_j \mid \mathcal{D}, Y_i=y_i}
\big[\ell(y_j,a)\big]
\right].
\]

The first term is the current expected loss on the validation set, and the second term is the expected loss after revealing the candidate’s label, updating the model, and then predicting optimally [2211.09283]. Since the first term is the same for all candidates, ranking candidates reduces to minimizing the expected post-query loss. In that sense, EER is a directly task-oriented acquisition rule rather than a generic uncertainty heuristic.

This formulation differs from criteria that score a point only by how uncertain the model is about that point itself. EER asks a counterfactual question about downstream performance: what reduction in future error is expected if this label is obtained now? This suggests that EER is intrinsically tied to a specified evaluation distribution, represented in the paper by the validation set [2211.09283].

## 2. Classical formulation and computational bottleneck

The original attraction of EER is that it targets the quantity of practical interest: expected future error. Its classical drawback is computational cost. In the original algorithm, one must, for each candidate \(x_i\), pretend that it has each possible label \(y_i\), retrain the model for each hypothetical label, and recompute the expected loss [2211.09283]. The paper states that this is feasible for very cheap models like Naive Bayes, but not for deep networks.

After dropping candidate-independent constants, the reduced selection score used in the Bayesian treatment is

\[
\text{Score}_i =
-
\mathbb{E}_{y_i \sim Y_i \mid \mathcal{D}}
\left[
\sum_{j=1}^{n_\text{val}}
\min_{a \in \mathcal{A}}
\mathbb{E}_{y_j \sim Y_j \mid \mathcal{D}, Y_i=y_i}
[\ell(y_j,a)]
\right].
\]

The computational bottleneck is therefore not conceptual but procedural: classical EER embeds repeated model updates inside the acquisition function itself [2211.09283]. This is the main reason it was “not been widely used for modern deep neural networks due to this large computational cost,” even though it had long been regarded as effective [2211.09283].

## 3. Bayesian reformulation for modern deep models

A recent reformulation places EER under the lens of Bayesian active learning and derives a computationally efficient version that can use any Bayesian parameter sampling method [2211.09283]. The key move is to replace explicit retraining with posterior predictive calculations. The derivation assumes a Bayesian posterior over parameters \(\theta\), conditional independence of labels given \(\theta\), and fixed unlabeled candidate and validation sets whose labels are treated as random variables [2211.09283].

With posterior samples \(\{\theta_t\}_{t=1}^T\), the required predictive quantities are approximated as

\[
\Pr(Y_i = c) \approx \frac{1}{T}\sum_{t=1}^T \Pr(Y_i=c \mid \theta_t),
\]

and

\[
\Pr(Y_i = c, Y_j = c') \approx \frac{1}{T}\sum_{t=1}^T
\Pr(Y_i=c \mid \theta_t)\Pr(Y_j=c' \mid \theta_t).
\]

The paper emphasizes that these marginals are sufficient for the Bayesian score, so no separate retraining is needed for each hypothetical label [2211.09283]. In practice, Monte Carlo dropout is used as the default approximate posterior sampler: dropout remains active at selection time, the model is run \(T\) times, and the resulting stochastic forward passes are treated as samples from \(\Pr(\theta\mid\mathcal{D})\) [2211.09283]. Cyclical SG-MCMC is also mentioned, but no clear empirical advantage over dropout is reported [2211.09283].

The resulting complexity is given as roughly

\[
O(K\, n_\text{pool}\, n_\text{val}\, T\, C^2),
\]

for MELL/MEZL, where \(K\) is the number of active learning rounds, \(T\) the number of posterior samples, and \(C\) the number of classes [2211.09283]. The paper notes that \(n_\text{pool}\) and \(n_\text{val}\) can be subsampled further.

## 4. Relation to uncertainty sampling and mutual information

The Bayesian reinterpretation also clarifies how EER differs from standard uncertainty-based acquisition. BALD chooses points that maximize mutual information between the candidate label and the model parameters,

\[
I(Y_i;\theta),
\]

which the paper describes as an epistemic-uncertainty criterion [2211.09283]. By contrast, the log-likelihood version of EER, MELL, is related to the mutual information between candidate labels and validation labels,

\[
\text{MELL} \sim \sum_{j=1}^{n_\text{val}} I(Y_i; Y_j),
\]

up to a constant [2211.09283].

The information-theoretic decomposition reported in the paper is

\[
H(Y_i) = I(Y_i;\theta) + H(Y_i \mid \theta),
\]

followed by

\[
I(Y_i;\theta) = I(Y_i;Y_j;\theta) + I(Y_i;\theta \mid Y_j) = I(Y_i;Y_j) + I(Y_i;\theta \mid Y_j).
\]

The interpretation given is that \(I(Y_i;Y_j)\) measures task-relevant information shared with the validation point, whereas \(I(Y_i;\theta \mid Y_j)\) captures information about parameters that is not relevant for predicting \(Y_j\) [2211.09283]. This is the basis for the claim that MELL can outperform BALD by focusing on information useful for the validation distribution rather than on parameter uncertainty in general.

A second distinction concerns loss choice. The paper reports that MEZL, the zero-one-loss variant, often underperforms MELL because a validation point contributes only if the candidate changes its MAP class prediction. If a query reduces uncertainty without flipping that prediction, MEZL gives it no credit [2211.09283]. MELL is therefore presented as empirically more reliable.

## 5. Empirical behavior, especially under dataset shift

The Bayesian EER framework is evaluated on seven image-classification datasets: CIFAR10, CIFAR100, SVHN, MNIST, Camelyon17, iWildCam, and FMoW, each in both no-shift and data-shift regimes, giving 14 experimental settings total [2211.09283]. For standard image datasets, shift is induced by splitting based on pixel brightness; for WILDS datasets, the shift is naturally present via predefined metadata splits [2211.09283].

The headline empirical result is that MELL is best or tied for best in 12 of 14 experiments, never does worse than BALD or Coreset in the reported AUC comparisons, and shows its biggest gains over BALD under data shift [2211.09283]. BADGE beats MELL on two shifted datasets, Camelyon17 and MNIST, but the paper notes that BADGE is an order of magnitude more expensive and that some runs did not complete after 14 days [2211.09283]. BADGE also requires access to last-layer embeddings, whereas MELL/MEZL are described as model-agnostic because they require only posterior predictive samples [2211.09283].

These findings support the paper’s practical claim that EER becomes especially attractive when the evaluation distribution differs from the seed distribution. Because the objective explicitly minimizes expected loss on a validation set, it is more aligned with target performance under shift than methods that measure epistemic uncertainty alone [2211.09283]. This suggests that the usefulness of EER is tied not only to query informativeness but also to how well the validation set represents the intended deployment distribution.

## 6. Terminological ambiguity across fields

The acronym “EER” does not have a uniform meaning across arXiv literatures.

| Usage of EER | Meaning | Example papers |
|---|---|---|
| Active learning | Expected Error Reduction | [2211.09283] |
| Biometrics, speaker verification, deepfake detection | Equal Error Rate | [1906.06262], [2104.00732], [2606.21584], [2309.12237], [2606.24586] |
| Factorial experiments | Experimentwise Error Rate | [2507.13621] |

In biometrics, EER is the equal error rate: the operating point at which false acceptance rate equals false rejection rate [1906.06262]. In speaker verification, the paper “Out of a hundred trials, how many errors does your speaker verifier make?” states explicitly that it is about Equal-Error Rate, not “Expected Error Reduction,” and treats EER as the point where \(miss(\theta^*) = fa(\theta^*) = EER\) [2104.00732]. In speech deepfake detection, EER is described as an oracle threshold metric computed on labeled test data, and the paper argues that it can hide deployment failure under domain shift [2606.21584]. In tandem biometric evaluation, t-EER extends equal error rate to a joint PAD-plus-comparator setting [2309.12237]. In deep biometric training, EERLoss is an approximation to equal error rate used as a loss function [2606.24586].

In replicated regular two-level factorial experiments, by contrast, EER means experimentwise error rate: the probability of making at least one false rejection among a family of hypotheses [2507.13621].

Because these meanings are all current in arXiv-indexed research, the abbreviation alone is potentially misleading. For active learning, the fuller term “Expected Error Reduction” avoids confusion.

## 7. Scope and practical interpretation

Within active learning, EER is best understood as a decision-theoretic acquisition principle: choose the label whose acquisition is expected to minimize future loss on a designated evaluation distribution [2211.09283]. Its historical weakness was retraining cost; its recent revival comes from a Bayesian posterior-sampling reformulation that avoids retraining for each candidate-label hypothesis [2211.09283].

The modern formulation also sharpens the conceptual distinction between EER and generic uncertainty sampling. BALD seeks information about parameters, whereas MELL seeks information relevant to validation labels [2211.09283]. This suggests that EER is most appropriate when the active-learning objective is not merely to reduce model uncertainty, but to reduce expected downstream error on a specific target distribution.

At the same time, the acronym’s cross-disciplinary ambiguity is substantial. In nearby fields, “EER” more often refers to equal error rate or experimentwise error rate than to Expected Error Reduction [1906.06262] [2104.00732] [2507.13621]. For that reason, technical writing typically benefits from spelling out the term on first use.

Source: https://www.emergentmind.com/topics/expected-error-reduction-eer