Active learning selects which unlabeled examples to label next so models learn efficiently. Traditional selection rules use prediction confidence or random sampling, but they can be misled by noisy data or overconfident models. Uncertainty-aware loss functions integrate the model’s own uncertainty into learning, creating a feedback loop that makes sample selection more informative and robust. Yarin Gal and Zoubin Ghahramani at the University of Cambridge showed how Bayesian approximations for deep models expose epistemic uncertainty, and Alex Kendall and Yarin Gal at the University of Cambridge clarified the distinction between aleatoric and epistemic uncertainty and how each should influence learning.
Mechanisms that link loss and selection
When the training objective explicitly includes an uncertainty term—for example by weighting the usual loss by predictive variance, adding a penalty for overconfidence, or optimizing mutual information—model gradients emphasize features that reduce epistemic uncertainty. Methods derived from Bayesian principles, such as the BALD (Bayesian Active Learning by Disagreement) criterion advocated in the literature by Gal and collaborators, select samples that maximize expected information gain. Practically, uncertainty-aware losses produce richer per-example signals: rather than treating every misprediction equally, the model highlights cases where additional labels would most reduce model uncertainty. This changes the sample ranking used by active learning acquisition functions and tends to prioritize borderline, out-of-distribution, or sparsely represented regions of the input space.
Relevance, causes, and consequences
Integrating uncertainty into loss functions is relevant because it aligns the training objective with the goal of reducing model ignorance, not merely minimizing immediate error. The cause is model miscalibration and limited data coverage; uncertainty-aware terms compensate by explicitly modeling what the model does not know. Consequences include fewer labels required to reach a target performance and improved generalization in under-sampled regions, which can reduce annotation cost and environmental impact from repeated retraining. There are important caveats: if uncertainty correlates with label noise or with underrepresented social groups, acquisition can inadvertently concentrate on noisy or sensitive samples, producing biased datasets. Human and cultural context matters because annotation difficulty, consent norms, and labeling cost vary territorially; practitioners should combine uncertainty-aware selection with domain expertise, fairness checks, and robust noise modeling to ensure ethically and scientifically sound data collection.