Ablation Based Counterfactuals

Diffusion models are a class of generative models that generate high-quality samples, but at present it is difficult to characterize how they depend upon their training data. This difficulty raises scientific and regulatory questions, and is a consequence of the complexity of diffusion models and their sampling process. To analyze this dependence, we introduce Ablation Based Counterfactuals (ABC), a method of performing counterfactual analysis that relies on model ablation rather than model retraining. In our approach, we train independent components of a model on different but overlapping splits of a training set. These components are then combined into a single model, from which the causal influence of any training sample can be removed by ablating a combination of model components. We demonstrate how we can construct a model like this using an ensemble of diffusion models. We then use this model to study the limits of training data attribution by enumerating full counterfactual landscapes, and show that single source attributability diminishes with increasing training data size. Finally, we demonstrate the existence of unattributable samples.

Training Data Attribution for Diffusion Models

Diffusion models have become increasingly popular for synthesizing high-quality samples based on training datasets. However, given the oftentimes enormous sizes of the training datasets, it is difficult to assess how training data impact the samples produced by a trained diffusion model. The difficulty of relating diffusion model inputs and outputs poses significant challenges to model explainability and training data attribution. Here we propose a novel solution that reveals how training data influence the output of diffusion models through the use of ensembles. In our approach individual models in an encoded ensemble are trained on carefully engineered splits of the overall training data to permit the identification of influential training examples. The resulting model ensembles enable efficient ablation of training data influence, allowing us to assess the impact of training data on model outputs. We demonstrate the viability of these ensembles as generative models and the validity of our approach to assessing influence.