The Journey, Not the Destination: How Data Guides Diffusion Models
Kristian Georgiev, Joshua Vendrow, Hadi Salman, Sung Min Park,, Aleksander Madry
TL;DR
This paper introduces a formal framework and efficient method for attributing generated images from diffusion models back to their training data, enabling better understanding and validation of data influence.
Contribution
It proposes a novel formal notion of data attribution for diffusion models and an efficient computational method to validate these attributions.
Findings
Successfully attributed images to training data in CIFAR-10 and MS COCO models.
Validated attributions through counterfactual analysis.
Provides open-source code for the attribution method.
Abstract
Diffusion models trained on large datasets can synthesize photo-realistic images of remarkable quality and diversity. However, attributing these images back to the training data-that is, identifying specific training examples which caused an image to be generated-remains a challenge. In this paper, we propose a framework that: (i) provides a formal notion of data attribution in the context of diffusion models, and (ii) allows us to counterfactually validate such attributions. Then, we provide a method for computing these attributions efficiently. Finally, we apply our method to find (and evaluate) such attributions for denoising diffusion probabilistic models trained on CIFAR-10 and latent diffusion models trained on MS COCO. We provide code at https://github.com/MadryLab/journey-TRAK .
Peer Reviews
Decision·Submitted to ICLR 2024
The choice of problem is a point of strength in this work. To understand the role of the training data in the final sampling result sheds light on the nature of the learned distribution and its properties. Also this question is tightly related to memorization and interpolation of training data which has practical implications such as data privacy.
The clarity of the writing can be significantly improved. The flow can be more streighforward. At its current form, the paper is too wordy to the extent that the main points are not clearly conveyed. Relying on a classifier to find emergence of features that are important throughout the diffusion trajectory is problematic. The sharp increase in classification performance under Fig3 might be due to this particular classifier and dataset. For example, for a more nuanced dataset with images with
1. The work focuses on an important and timely problem. Attribution for models is an important technical problem to address as generative models become more ubiquitous. This has implications for regulation, copyright, and fair compensation to artists [1]. 2. The proposed solution is reasonably motivated and builds on prior work that achieves SOTA attribution results for discriminative models. The work also compares against reasonable baselines for data attribution and shows results across two di
1. A limitation of the proposed approach is the fact that attribution scores are only provided for a single denoising step. This is unintuitive, as it requires multiple steps to be analyzed to understand how an image was generated. It would be good to obtain a single-shot attribution score for the entire diffusion trajectory. While simple heuristics can be employed to obtain this from the current approach, it's unclear if these are useful and interpretable. 2. There is little analysis regarding
- This submission tackles an interesting and important problem that to my knowledge was not yet approached in the context of diffusion models - The proposed method is a direct application of the Trek method to diffusion models. However, this extension is non-trivial and might have significant impact on some areas of research such as machine unlearning. - The evaluation is performed on a big scale (although it is mostly presented in the appendices)
-In general, there is a significant mismatch between the goal of the method as expressed at the beginning of the submission and the final experiments being evaluated. This is due to a series of approximations that make the computation possible. The interesting question is how those approximations influence the final observations (e.g. using one-step approximation of $\hat{x}_0^t$ as an approximation of the distribution’s expectation $E[x_0|x_t]$). If I understand it correctly, this assumption in
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsGenerative Adversarial Networks and Image Synthesis · Advanced Neuroimaging Techniques and Applications · AI in cancer detection
MethodsDiffusion