Robust Semisupervised Deep Generative Model Under Compound Noise

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Semisupervised learning is crucial for deep generative models like variational autoencoders.
• Existing models struggle with compound noise, simultaneously affecting data and labels (outliers and noisy labels).

Purpose of the Study:

• To propose a novel noise-robust semisupervised deep generative model.
• To jointly address noisy labels and outliers in a unified framework.

Main Methods:

• Developed a unified robust semisupervised variational autoencoder (URSVAE) considering input data uncertainty.
• Integrated a denoising layer into URSVAE for label correction.
• Employed robust beta-divergence for variational inference robustness against outliers.
• Combined URSVAE decoder with a generative adversarial network generator into URSVAE-GAN.

Main Results:

• The URSVAE-GAN framework effectively handles compound noise (outliers and noisy labels).
• Achieved competitive performance using a novel variational lower bound derived from beta-divergence.
• Demonstrated superior performance on image classification and face recognition tasks.

Conclusions:

• The proposed URSVAE-GAN framework offers a robust solution for semisupervised deep generative modeling with compound noise.
• The end-to-end denoising scheme in joint optimization enhances model performance.
• The framework outperforms state-of-the-art approaches in relevant benchmarks.