Probability Density from Latent Diffusion Models for Out-of-Distribution Detection

Likelihood-based detection works better in representation space than pixel space: The research found that using data likelihood as an out-of-distribution detector performs significantly better when applied to learned feature representations (from ResNet-18) rather than raw images, suggesting previous failures of likelihood-based methods may have been due to technical limitations rather than fundamental flaws in the approach.

Theoretical optimality of likelihood under standard assumptions: The paper demonstrates mathematically that likelihood is the theoretically optimal out-of-distribution detector when OOD data is uniformly distributed, providing theoretical justification for pursuing likelihood-based approaches despite earlier disappointing practical results.

Representation learning is key to safety: The findings highlight that how AI systems learn to represent data internally—not just the raw input data itself—is crucial for building reliable safety mechanisms like OOD detection, suggesting that investment in better representation learning could improve AI system robustness.

Practical path forward for deployment safety: By showing that diffusion models trained on learned representations can match or exceed state-of-the-art detection methods, this work offers a concrete, implementable approach for organizations looking to add out-of-distribution detection to their machine learning systems before deployment.