Optimal ablation for interpretability
This work addresses interpretability challenges for researchers and practitioners in machine learning, offering an incremental improvement over existing ablation techniques.
The paper tackles the problem of quantifying the importance of model components for interpretability by proposing optimal ablation (OA), which shows theoretical and empirical advantages over prior ablation methods and benefits tasks like circuit discovery and factual recall localization.
Interpretability studies often involve tracing the flow of information through machine learning models to identify specific model components that perform relevant computations for tasks of interest. Prior work quantifies the importance of a model component on a particular task by measuring the impact of performing ablation on that component, or simulating model inference with the component disabled. We propose a new method, optimal ablation (OA), and show that OA-based component importance has theoretical and empirical advantages over measuring importance via other ablation methods. We also show that OA-based component importance can benefit several downstream interpretability tasks, including circuit discovery, localization of factual recall, and latent prediction.