Generative-Discriminative Complementary Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• State-of-the-art deep learning often uses discriminative models requiring high-quality labeled data.
• Obtaining such data is challenging, especially with numerous classes.
• Complementary labels offer a simpler annotation process (yes/no per class).

Purpose of the Study:

• To address the challenge of data scarcity in deep learning by utilizing complementary labels.
• To propose a novel method that combines generative and discriminative approaches for complementary learning.
• To improve the accuracy of predicting ordinary labels and generate high-quality instances under weak supervision.

Main Methods:

• We propose Complementary Conditional GAN (CCGAN), a generative-discriminative model.
• CCGAN models both conditional (discriminative) and instance (generative) distributions.
• The method estimates ordinary labels from complementary labels.

Main Results:

• CCGAN significantly improves the accuracy of predicting ordinary labels.
• The model demonstrates the capability to generate high-quality instances.
• Extensive empirical studies validate the effectiveness of CCGAN.
• Theoretical analysis confirms the model's ability to retrieve true conditional distributions.

Conclusions:

• Complementary Conditional GAN (CCGAN) offers an effective solution for deep learning with complementary labels.
• The proposed method enhances label prediction accuracy and data generation capabilities.
• CCGAN provides a robust framework for learning from weakly supervised data.