Generative Classifiers Outperform Discriminative Ones on Distribution Shift

This paper addresses a critical problem in machine learning: the vulnerability of discriminative classifiers to distribution shifts due to their reliance on spurious correlations. It proposes and demonstrates the effectiveness of generative classifiers as a more robust alternative. The paper's significance lies in its potential to improve the reliability and generalizability of AI models, especially in real-world applications where data distributions can vary.

• •Discriminative classifiers often fail under distribution shift due to reliance on spurious correlations.
• •Generative classifiers, using class-conditional generative models, are proposed as a more robust alternative.
• •Diffusion-based and autoregressive generative classifiers achieve state-of-the-art performance on distribution shift benchmarks.
• •Generative classifiers reduce the impact of spurious correlations in realistic applications.
• •The paper provides analysis of generative classifier inductive biases and data properties for optimal performance.