Introduction: The Challenge of Instance-dependent Partial-label Learning
Partial-label learning (PLL) is a framework where each training instance is associated with a set of candidate labels, only one of which is the ground truth. Traditional PLL research often assumes that the "extra" incorrect labels are chosen at random (instance-independent). However, in real-world scenarios, labeling errors are rarely random; an annotator is far more likely to mistake a "Huskie" for a "Wolf" than for a "Whale."
This phenomenon is known as Instance-dependent PLL (ID-PLL). The primary challenge here is under-disambiguation (UD). Because incorrect candidate labels are often semantically similar to the true label, models struggle to distinguish between them, resulting in "flat" prediction probabilities across the candidate set. Standard PLL methods typically fail in these cases because they force the model to commit to a single label too early, leading to noisy supervision and degraded performance.
Methodology: Selective Disambiguation via Normalized Entropy
To address the UD problem, this paper proposes a two-stage framework centered on Normalized Entropy (NE). NE serves as a metric to gauge how effectively a model has disambiguated an instance. Unlike standard entropy, NE accounts for the varying number of candidate labels across different examples, providing a fairer basis for comparison.
Stage 1: Selective Training
The framework categorizes training examples into two groups based on their NE:
- Well-Disambiguated (WD) Examples: Instances with low NE, where the model shows high confidence in a specific candidate. These receive strong supervision.
- Under-Disambiguated (UD) Examples: Instances where the model remains uncertain. Instead of forcing a guess, the model utilizes complementary supervision, learning from labels known to be false (those outside the candidate set). This provides "harmless" information that aids convergence without introducing noise.
Stage 2: Candidate-aware Thresholding (CAT)
Recognizing that some classes are inherently harder to learn than others, the authors implement Candidate-aware Thresholding (CAT). This involves class-wise WD-NE thresholds rather than a single global value, preventing easy-to-learn classes from dominating the training process. Additionally, a dual-network agreement strategy is employed: an example is only promoted to the WD set if both networks agree on the predicted label and the NE is sufficiently low.
Experimental Results
The proposed method was evaluated on several benchmarks, including CIFAR-10 and CIFAR-100, using simulated ID-PLL labels. The results demonstrate a significant performance leap over existing state-of-the-art (SOTA) methods:
- CIFAR-10: The method achieved 95.04% accuracy, outperforming top competitors like ABLE (92.25%) and PiCO (91.15%).
- CIFAR-100: It reached 78.67%, a nearly 4% improvement over the previous best, CRDPLL (77.82%).
The framework proved remarkably robust, maintaining stable performance even as the degree of label ambiguity ($q$) increased, a scenario where traditional PLL methods typically see sharp declines.
Conclusion
By using Normalized Entropy as a "progress bar" for learning, this framework effectively navigates the complexities of instance-dependent noise. The combination of selective disambiguation and class-wise thresholding allows the model to leverage high-confidence predictions while safely extracting information from ambiguous examples through complementary labels. While the dual-network approach adds some computational overhead, the substantial gains in accuracy and robustness make it a powerful solution for real-world partial-label learning tasks.
Summary updated