Quantifying explainability for unsupervised representation learning requires operational measures that connect latent structure to human-understandable concepts, downstream utility, and robustness. Foundational guidance from Yoshua Bengio at Université de Montréal emphasizes that representations should be useful, disentangled, and robust. Practical quantification therefore combines predictive proxies, information-theoretic scores, and task-based evaluations.
Quantitative measures
Reconstruction error and likelihood estimate model fidelity but do not measure interpretability directly. Predictive probes that train a simple classifier or regressor on latent codes to predict known generative factors or labels make informativeness measurable through accuracy or explained variance. Clustering quality metrics such as adjusted rand index and silhouette score assess whether latent groupings align with meaningful categories. Information-theoretic quantities including mutual information and total correlation quantify redundancy and independence among latent dimensions, and reduced total correlation often correlates with disentanglement in practice. Work on constrained variational autoencoders led by Irina Higgins at DeepMind demonstrated that encouraging factorized latents improves alignment between individual latent dimensions and semantic factors, making such statistics actionable.
Causes and consequences
Causes of poor explainability include model capacity that mixes factors, training objectives that only optimize reconstruction, and unobserved confounders in the data. These lead to latent features that are brittle under distribution shift, opaque to domain experts, and potentially biased when used in decision pipelines. From a societal and regulatory perspective, opaque unsupervised features can amplify harms in healthcare or criminal justice settings where provenance and interpretability matter. Pieter Abbeel at UC Berkeley and colleagues emphasize evaluating representations by downstream transfer performance as a pragmatic proxy for usefulness in real-world tasks, linking quantification directly to consequences for deployment.
Measuring explainability thus benefits from multi-faceted protocols that report several metrics together. Reporting predictive probe performance, disentanglement proxies like independence measures, robustness under perturbations, and downstream task transfer paints a fuller picture than any single number. Nuance is important: high disentanglement metrics do not guarantee human interpretability if the dataset lacks labeled factors, and strong downstream performance can coexist with opaque latent semantics. Combining these quantitative measures with domain expert inspection and provenance tracking makes explainability both measurable and actionable for unsupervised representation learning.