UASPL: Uncertainty-Aware Self-Paced Learning with Evidential Neural Networks

Technical Summary: Uncertainty-Aware Self-Paced Learning with Evidential Neural Networks (UASPL)

Problem Statement
Self-paced learning (SPL) is a paradigm that mimics human learning by training models on samples ranging from easy to difficult, typically determined by the magnitude of the training loss. While effective, the authors identify a critical flaw in traditional SPL: samples with low loss values in early training stages are not necessarily "reliably easy" samples. Experimental analysis reveals that the Mean Relative Loss Variation (MRLV) of samples initially selected as "easy" often fluctuates or increases significantly in subsequent training rounds. This indicates that low-loss samples may be pseudo-easy—either lacking sufficient evidence for the model to learn effectively or representing misleading predictions that the model is overconfident about. Consequently, relying solely on loss-based criteria fails to distinguish truly simple samples from those that are merely easy to fit temporarily but unreliable.

Methodology
To address this, the paper proposes UASPL (Uncertainty-Aware Self-Paced Learning), a framework that integrates Evidential Neural Networks (ENN) within the Subjective Logic framework. Unlike traditional methods that rely on point-estimate probabilities, UASPL utilizes a Dirichlet distribution to parameterize class probabilities, allowing for the direct quantification of predictive uncertainty.

The core of UASPL is a novel uncertainty-aware loss function that combines the label-fitting loss with an uncertainty-weighted regularization term. The total loss for the $i$-th sample, $L^{(i)}_{total}$, is formulated as:
$$L^{(i)}_{total} = L^{(i)}_{EMSE} + \text{coeff}_i \cdot L^{(i)}_{KL}$$

Where:

• $L^{(i)}_{EMSE}$ is the Evidential Mean Square Error, measuring the fit between predicted probabilities and ground truth.
• $L^{(i)}_{KL}$ is a Kullback–Leibler divergence term that penalizes evidence accumulation on non-ground-truth classes.
• $\text{coeff}_i$ is an adaptive coefficient defined as $(1-c_i)(1-u_i) + c_i u_i$, where $c_i$ is the prediction correctness indicator (1 if correct, 0 otherwise) and $u_i$ is the evidential uncertainty.

This design creates a dynamic selection preference:

1. Correct predictions with low uncertainty: These are identified as reliably simple samples and are prioritized.
2. Correct predictions with high uncertainty: These are treated as samples within the learning scope but lacking sufficient evidence; the KL term is strengthened to guide reliable evidence growth.
3. Incorrect predictions with high uncertainty: The model is allowed to explore these samples more freely as the lack of evidence suggests the prediction is not yet confident.
4. Incorrect predictions with low uncertainty: These represent confident but misleading errors. The KL term is amplified to suppress the accumulation of erroneous evidence.

The algorithm iteratively updates model parameters and sample weights ($v_i$) based on this composite loss, effectively embedding uncertainty estimation directly into the sample selection criterion without requiring external priors.

Key Contributions

1. First Integration of Evidential Uncertainty in SPL: UASPL is the first self-paced learning method to directly incorporate model-generated evidential uncertainty alongside label-fitting loss into the optimization objective. This ensures the selection of samples that are not only low-loss but also reliable.
2. Interpretability: The paper provides a theoretical analysis demonstrating that the sample-selection preference induced by UASPL aligns with the "easy-to-hard" principle. It explicitly prioritizes reliable easy samples while deferring those with insufficient or misleading evidence.
3. Generality: The proposed loss function is designed as a general framework that can be extended to different SPL variants (e.g., linear or mixture regularizers) without substantial modification, proving it is not tied to a specific regularization form.

Experimental Results
The authors evaluated UASPL on multiple datasets, including 25 tabular datasets from the UCI repository (both balanced and imbalanced) and four image classification benchmarks (CIFAR-10, FashionMNIST, MNIST, SVHN). The method was compared against 15 baselines, including traditional SPL variants, methods incorporating external uncertainty priors, and adapted SPL methods using various uncertainty criteria (e.g., confidence, margin, MC dropout).

The results indicate that UASPL outperforms other SPL methods in terms of:

• Classification Performance: Achieving higher accuracy and better stability across datasets.
• Interpretability: Successfully selecting samples that remain stable in difficulty across training rounds, unlike traditional loss-based selection.
• Generality: Demonstrating effectiveness across different data modalities and SPL formulations.

Significance
The paper claims that UASPL resolves the fundamental issue of "pseudo-easy" samples in self-paced learning by shifting the selection criterion from purely loss-based to uncertainty-aware. By leveraging the internal uncertainty estimates of evidential neural networks, UASPL ensures that the curriculum learning process focuses on samples that the model can reliably learn from, thereby improving training efficiency and final model robustness. The authors position this as a significant step forward in making SPL more theoretically sound and practically effective, particularly in scenarios where model confidence and evidence sufficiency are critical.