Local and Global Guidance for Multi-Complementary Label Learning
Cheng Chen, Ivor Wai-Hung Tsang · 2024
Complementary label learning involves instances associated with a candidate set of labels, which may contain multiple labels or a label that does not belong to the true label. Even though such weak supervision is easy to obtain, the setup inevitably introduces issues of label ambiguity and potential label distribution bias. Previous works have approached the task either based on the local characteristics of the dataset (label generation-wise), improving model robustness through mechanisms like transition matrices, or leveraging the global characteristics of the dataset (instance-wise), such as data augmentation and sample selection. However, these methods fail to simultaneously exploit both local and global aspects in addressing the issues of the multi-complementary label task. To tackle this issue, our paper presents a unified approach that considers both local and global guidance. For local guidance, we design homogeneous label regularisation using optimal transportation via the Sinkhorn-Knopp algorithm, ensuring a uniform label distribution across instances. For global guidance, we implement co-occurrence class embedding regularisation to enhance the model’s understanding of the underlying sample distribution. This dual approach mitigates the misleading effects of complementary labels and corrects errors introduced by less informative labels during training. Overall, our experiments have demonstrated that our method is particularly effective on large datasets like CIFAR-100.