Improving Causal Interventions in Amnesic Probing with Mean Projection or LEACE

2025-06-13

AI Summary

This study addresses the problem of impure target-information erasure in amnesic probing, which leads to distorted causal attribution. We propose two novel linear erasure paradigms—Mean Projection (MP) and LEACE (Linear Erasure of Associated Concept Embeddings)—featuring stronger geometric constraints and reduced semantic interference. We first systematically expose the inherent bias in INLP arising from random perturbations, then enhance erasure purity via strict orthogonalization and covariance calibration. Experiments across multiple language tasks show that MP and LEACE improve target-information removal by 23–41%, preserve non-target features at 98.7%, and double the accuracy of behavioral-change attribution—thereby significantly improving the causal fidelity of representation interventions.

Technology Category

• Machine Learning: Causal Learning
• Knowledge Representation and Reasoning: Action, Change, and Causality
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Security and Privacy: Data transparency and provenance

Abstract

Amnesic probing is a technique used to examine the influence of specific linguistic information on the behaviour of a model. This involves identifying and removing the relevant information and then assessing whether the model's performance on the main task changes. If the removed information is relevant, the model's performance should decline. The difficulty with this approach lies in removing only the target information while leaving other information unchanged. It has been shown that Iterative Nullspace Projection (INLP), a widely used removal technique, introduces random modifications to representations when eliminating target information. We demonstrate that Mean Projection (MP) and LEACE, two proposed alternatives, remove information in a more targeted manner, thereby enhancing the potential for obtaining behavioural explanations through Amnesic Probing.

Problem

• Targeted removal of specific linguistic information in models
• Avoiding random modifications during information elimination
• Enhancing behavioral explanations via improved amnesic probing techniques

Innovation

• Uses Mean Projection for targeted information removal
• Employs LEACE to enhance behavioral explanations
• Improves amnesic probing by minimizing random modifications