Deeper Diffusion Models Amplify Bias

Despite the impressive performance of generative Diffusion Models (DMs), their internal working is still not well understood, which is potentially problematic. This paper focuses on exploring the important notion of bias-variance tradeoff in diffusion models. Providing a systematic foundation for this exploration, it establishes that at one extreme the diffusion models may amplify the inherent bias in the training data and, on the other, they may compromise the presumed privacy of the training samples. Our exploration aligns with the memorization-generalization understanding of the generative models, but it also expands further along this spectrum beyond ``generalization'', revealing the risk of bias amplification in deeper models. Building on the insights, we also introduce a training-free method to improve output quality in text-to-image and image-to-image generation. By progressively encouraging temporary high variance in the generation process with partial bypassing of the mid-block's contribution in the denoising process of DMs, our method consistently improves generative image quality with zero training cost. Our claims are validated both theoretically and empirically.

Post-hoc Spurious Correlation Neutralization with Single-Weight Fictitious Class Unlearning

Neural network training tends to exploit the simplest features as shortcuts to greedily minimize training loss. However, some of these features might be spuriously correlated with the target labels, leading to incorrect predictions by the model. Several methods have been proposed to address this issue. Focusing on suppressing the spurious correlations with model training, they not only incur additional training cost, but also have limited practical utility as the model misbehavior due to spurious relations is usually discovered after its deployment. It is also often overlooked that spuriousness is a subjective notion. Hence, the precise questions that must be investigated are; to what degree a feature is spurious, and how we can proportionally distract the model's attention from it for reliable prediction. To this end, we propose a method that enables post-hoc neutralization of spurious feature impact, controllable to an arbitrary degree. We conceptualize spurious features as fictitious sub-classes within the original classes, which can be eliminated by a class removal scheme. We then propose a unique precise class removal technique that employs a single-weight modification, which entails negligible performance compromise for the remaining classes. We perform extensive experiments, demonstrating that by editing just a single weight in a post-hoc manner, our method achieves highly competitive, or better performance against the state-of-the-art methods.