UASPL: new methodology for self-learning of neural networks takes uncertainty of predictions into account

Researchers in July 2026 described in a paper on arXiv the UASPL method (Uncertainty-Aware Self-Paced Learning) — a modification of self-paced learning that considers not only the magnitude of model error but also the reliability of its predictions, using evidential neural networks and the Subjective Logic framework.

What's the problem with classic self-paced learning

Self-paced learning (SPL) is a popular approach to neural network training that mimics the human learning process: the network first "learns" simple examples, then gradually moves to complex ones, guided by the loss function value on each sample. The logic is straightforward: the smaller the model's error on an example, the "simpler" it's considered, and such examples are included in training first.

UASPL's authors point out a weak spot in this logic: small loss doesn't always mean the model's prediction is truly reliable. A model might give low error on an example that is actually complex or contains noise — and then the algorithm mistakenly classifies it as "easy" and trains on it earlier than it should, potentially cementing an erroneous pattern.

How UASPL works

To solve this problem, the authors proposed embedding an estimate of prediction uncertainty into example selection through evidential neural networks — an architecture that outputs not just the prediction itself but also an estimate of confidence in it. This estimate is formalized within the Subjective Logic framework — a mathematical apparatus for describing unreliable judgments, where each statement is accompanied by a degree of confidence.

This approach is part of a broader evidential deep learning trend: instead of outputting a single probability number for a class, as in standard softmax classification, the network learns to assess "confidence in its own confidence." In practice, this allows distinguishing between situations where a model genuinely understands an example and situations where low loss occurred by chance — for instance, due to noise in annotation or atypical feature distribution in a specific sample.

• The method is based on a general loss function applicable to various self-paced learning variants, not just one specific implementation
• The loss function simultaneously addresses two tasks: estimating uncertainty and setting preference when selecting the next training examples
• This pairing provides interpretability: it's clear why a specific example was selected or deferred
• The method was tested on multiple datasets for classification tasks
• The source code is released publicly on GitHub (repository treelife979/UASPL)

What experiments revealed

Based on testing across multiple datasets, UASPL outperformed other self-paced learning methods in three areas: classification accuracy, interpretability of the sample selection process, and generality — the ability to maintain effectiveness when transferring to different tasks and datasets, rather than just the configuration initially tuned for.

What this means

For teams training models on real, imperfectly labeled data, UASPL offers a more cautious curriculum learning principle: not "simplicity first" but "reliability first." This can reduce the risk that the model will cement erroneous patterns early in training due to noisy, but formally "easy" examples — which is especially relevant for tasks where data is collected automatically and doesn't undergo manual review, such as medical image annotation, user content moderation, or sensor data collection from industrial systems.

The fact that the method is described in the paper as a "general loss function," extensible to different variants of self-paced learning, matters practically: teams don't need to rewrite their entire training pipeline from scratch; they can adapt an existing SPL scheme by adding uncertainty estimation. Open source code on GitHub lowers the entry barrier for testing the method on own datasets — this is particularly valuable for highly specialized research topics where ready-made industrial implementations don't yet exist.