UASPL: a new self-paced training method that accounts for prediction uncertainty

In July 2026, researchers described UASPL (Uncertainty-Aware Self-Paced Learning) in a paper on arXiv. It modifies self-paced learning so that it accounts not only for the size of the model's error but also for the reliability of its predictions, using evidential neural networks and the Subjective Logic framework.

What is wrong with classic self-paced learning

Self-paced learning (SPL) is a popular approach to training neural networks that mimics how humans learn: the network first "studies" easy examples and then gradually moves on to harder ones, guided by the value of the loss function on each sample. The logic is simple: the smaller the model's error on an example, the "easier" it is considered, and such examples enter training earlier.

The UASPL authors point to a weakness in this logic: a small loss does not always mean the model's prediction is actually reliable. A model may report an understated error on an example that is in fact difficult or noisy. The algorithm then wrongly counts it as "easy" and trains on it earlier than it should, potentially entrenching an erroneous pattern.

How UASPL works

To address this, the authors built an estimate of prediction uncertainty into example selection using evidential neural networks, an architecture that outputs a confidence estimate alongside the prediction itself. This estimate is formalized within the Subjective Logic framework, a mathematical apparatus for describing uncertain judgments in which every statement carries a degree of trust.

The approach is part of the broader field of evidential deep learning: instead of producing a single class probability, as in ordinary softmax classification, the network also learns to estimate its "confidence in its own confidence." In practice, this makes it possible to tell apart a case where the model really understands an example from one where a low loss came about by chance, for instance because of label noise or an atypical distribution of features in a particular sample.

• The method rests on a general loss function that applies to different variants of self-paced learning, not just one specific implementation
• The loss function does two things at once: it estimates uncertainty and sets the preference for choosing the next training examples
• This combination makes the method interpretable: it is visible why a given example was selected or postponed
• The method was tested on several datasets in classification tasks
• The source code is publicly available on GitHub (repository treelife979/UASPL)

What the experiments showed

In tests on several datasets, UASPL outperformed other self-paced learning methods in three areas at once: classification accuracy, interpretability of the sample selection process, and generality, meaning the ability to stay effective when transferred to different tasks and datasets, rather than only in the configuration the method was originally tuned for.

What this means

For teams that train models on real, imperfectly labeled data, UASPL offers a more cautious curriculum learning principle: not "easy first" but "reliable first." This may reduce the risk that a model locks in erroneous patterns early in training because of noisy but formally "easy" examples. That matters most for tasks where data is collected automatically and not manually checked, such as labeling medical images, moderating user content or gathering sensor data from industrial systems.

The fact that the paper describes the method as a "general loss function" extensible to different variants of self-paced learning matters in practice: teams may not need to rewrite the whole training pipeline from scratch, and can instead adapt an existing SPL scheme by adding an uncertainty estimate. The open source code on GitHub lowers the barrier to testing the method on one's own datasets. This is especially valuable for narrowly specialized research topics, where ready-made industrial implementations do not yet exist.