GAN-CL: Generative Adversarial Networks for Learning From Complementary Labels.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Complementary labels (CLs) offer a less laborious alternative to traditional supervised learning labels.
• However, the weaker signal from CLs often hinders effective feature representation learning and degrades classifier performance.

Purpose of the Study:

• To develop an effective algorithm for learning from complementary labels.
• To leverage generative adversarial networks (GANs) to enhance feature representation and classifier performance in the CL setting.

Main Methods:

• Introduced GAN-CL, an algorithm that utilizes GANs for complementary label learning.
• Adapted the discriminator in GAN-CL to function as a classifier, incorporating complementary information into its objective.
• Analyzed the global optimality of the generator and discriminator within the GAN-CL framework.

Main Results:

• Demonstrated significant improvements in learning from CLs using the proposed GAN-CL algorithm.
• Achieved compelling performance gains compared to state-of-the-art complementary label learning approaches.
• Validated the effectiveness of GAN-CL through extensive experiments on benchmark image datasets using deep models.

Conclusions:

• GAN-CL effectively addresses the challenge of learning from less informative complementary labels.
• The proposed method offers a robust solution for improving classifier performance in complementary supervision scenarios.
• The findings highlight the potential of GANs in advancing machine learning paradigms with weaker supervision signals.