OpenGAN: Open-Set Recognition via Open Data Generation.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Real-world machine learning requires analyzing data that differs from training data, a challenge known as open-set recognition.
• Existing methods like binary discriminators or GAN discriminators have limitations in generalizing to diverse open-set data or suffer from unstable training.

Purpose of the Study:

• To develop an improved method for open-set recognition that effectively discriminates between known (closed-set) and unknown (open-set) data.
• To address the generalization issues of binary discriminators and the instability of GANs in open-set scenarios.

Main Methods:

• Proposed OpenGAN, combining GANs with existing K-way classifiers.
• Augmented real outlier data with adversarially synthesized 'fake' data for training.
• Built the discriminator on features from closed-world K-way networks, enabling a lightweight discriminator head.

Main Results:

• Demonstrated that a carefully selected GAN-discriminator on real outlier data achieves state-of-the-art performance.
• Showcased OpenGAN's ability to significantly outperform prior open-set recognition methods.
• Validated the effectiveness of augmenting training data with synthesized outliers.

Conclusions:

• OpenGAN offers a robust and efficient solution for open-set recognition by leveraging feature representations from established classifiers.
• The method effectively handles diverse open-test data by mitigating overfitting to specific outliers and improving GAN stability.
• OpenGAN represents a significant advancement in machine learning systems requiring robust analysis of unfamiliar data patterns.