A new research paper challenges the common practice of using marginal matching to verify independence between style variables and class information in factorized generative models. The authors demonstrate that matching only the marginal distribution is insufficient, as it allows the style variable to remain highly predictive of the class label. Their study shows that even models with near-zero global MMD can still leak class information, with linear probes achieving high accuracy in recovering labels. The paper proposes mitigation strategies that reduce probe accuracy but do not fully eliminate within-class dependence, highlighting the limitations of relying solely on marginal statistics for certifying independence. AI
IMPACT Highlights a potential flaw in the evaluation of generative models, impacting how their style and content independence is assessed.
RANK_REASON Academic paper detailing a theoretical finding and empirical validation. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
