Mixed Blessing: Class-Wise Embedding guided Instance-Dependent Partial Label Learning

TL;DR

This work tackles instance-dependent partial label learning (IDPLL), where candidate labels are highly correlated with sample features, offering both informative supervision and high label ambiguity. It introduces a three-module framework that learns per-sample class-wise embeddings, enabling explicit modeling of relationships among class labels. Two losses—class associative loss $L_{cal}$ and prototype discriminative loss $L_{pdl}$—drive embeddings to cluster within candidate labels while aligning high-confidence predictions with global class prototypes. Training is staged: first optimize $\,\mathcal{L}_{cls} + \alpha \, \mathcal{L}_{cal}$, then incorporate $eta \, \mathcal{L}_{pdl}$ to form $\\mathcal{L}_{all} = \, \mathcal{L}_{cls} + \alpha \, \mathcal{L}_{cal} + \beta \, \mathcal{L}_{pdl}$. Experiments on six benchmarks show consistent improvements over twelve baselines, with particular gains on fine-grained datasets and faster early learning, and code is publicly available at https://github.com/Yangfc-ML/CEL.

Abstract

In partial label learning (PLL), every sample is associated with a candidate label set comprising the ground-truth label and several noisy labels. The conventional PLL assumes the noisy labels are randomly generated (instance-independent), while in practical scenarios, the noisy labels are always instance-dependent and are highly related to the sample features, leading to the instance-dependent partial label learning (IDPLL) problem. Instance-dependent noisy label is a double-edged sword. On one side, it may promote model training as the noisy labels can depict the sample to some extent. On the other side, it brings high label ambiguity as the noisy labels are quite undistinguishable from the ground-truth label. To leverage the nuances of IDPLL effectively, for the first time we create class-wise embeddings for each sample, which allow us to explore the relationship of instance-dependent noisy labels, i.e., the class-wise embeddings in the candidate label set should have high similarity, while the class-wise embeddings between the candidate label set and the non-candidate label set should have high dissimilarity. Moreover, to reduce the high label ambiguity, we introduce the concept of class prototypes containing global feature information to disambiguate the candidate label set. Extensive experimental comparisons with twelve methods on six benchmark data sets, including four fine-grained data sets, demonstrate the effectiveness of the proposed method. The code implementation is publicly available at https://github.com/Yangfc-ML/CEL.

Figures (6)

• Figure 1: Differences between the conventional PLL and IDPLL, where the red label is the ground-truth label of the instance. In PLL, the noisy labels in the candidate label set are randomly generated. However, in IDPLL, the noisy labels in the candidate label set are instance-dependent, making them very similar to that of the ground-truth label, which brings more label ambiguity.
• Figure 2: The classification accuracy curves of PLL method PRODEN PRODEN in IDPLL and PLL settings on two data sets CUB200 and FGVC100, where AVG. CLs represent the average number of candidate labels of each sample. In the IDPLL setting, the model has a faster learning speed in the early stage of training because the instance-dependent noisy labels are related to the sample to some extent, which can provide more supervision in the early stage of model training. However, in the later stage of training, the performance of PRODEN in the IDPLL setting is significantly inferior to that in the PLL setting as instance-dependent noisy labels bring more label ambiguity.
• Figure 3: Illustration of our method CEL. Our model consists of three modules: the backbone $f$, the class-wise encoder $g$, and the classifier $z$. For each sample, after processing through the backbone $f$ and class-wise encoder $g$, each sample obtains its class-wise embeddings, i.e., each class corresponds to an embedding for that sample. The red line represents the process of constructing class prototypes based on the high-confidence class selected according to the model's output. The class associative loss (CAL) considers the relationships among each sample's different class-wise embeddings. While the prototype discriminative loss (PDL) considers the relationships between high-confidence class and class prototypes.
• Figure 4: Classification accuracy curves of all methods on benchmark data set CUB200.
• Figure 5: Parameters sensitivity of our method CEL. (a) - (d) represent the classification accuracy of our method on benchmark data sets CIFAR-100 and CUB200 by varying $\alpha$, $\beta$, $\gamma_1$ and $\gamma_2$ respectively.