Disentangling Influence: Using Disentangled Representations to Audit Model Predictions
summary
Method reduces indirect influence auditing to direct influence by learning a disentangled latent where protected factor is independent, then composing its decoder with the original model and measuring direct influence. Implementation uses an autoencoder with adversary to hide protected info, computes direct attributions with SHAP, and reports per-instance indirect influence plus reconstruction and prediction-error flags. Experiments on synthetic generative data, pixelized shapes, and Adult income show the method finds proxy paths and latent influences but depends on brittle disentanglement and imperfect reconstructions.
AI-written summary · about
chapters (23)
- 0:00Introduction
- 0:22Background
- 2:10General Question
- 2:23Direct and Indirect Influence
- 3:36Goal
- 5:40Driving Methodology
- 6:18Independent Settings
- 6:41Independent Factors
- 7:13Group Actions
- 8:25Independent World States
- 9:30Indirect Influence
- 11:38Supervised Learning
- 12:00Combining Processes
- 12:22Independent Features
- 12:42Autoencoder
- 14:04Experimental Results
- 17:14Reconstruction Error
- 17:35Prediction Error
- 18:40Feature Audit
- 19:47Shape Scale Orientation
- 23:04Recurse
- 24:48Limitations
- 25:55Questions
related talks
46mAI for High-Stakes Decision Making
28mLearning Causal Representations From Unknown Interventions
51mEfficient Retrieval of Influential LLM Training Examples
48mUnveiling the secrets of a machine learning model
43mToward Human-Centered Machine Learning
21mThe Emperor's New Embeddings: Obfuscating ML Inputs Doesn't Provide Privacy
15mInterpretable Machine Learning with R
25mWhy No Model Is A Black Box