A Two-Way Survey of Intrinsically Interpretable Deep Learning for Tabular Data: Bridging Methods and Systems

The interpretability crisis in Artificial Intelligence (AI) has intensified with the widespread adoption of complex closed box models in critical domains such as healthcare, finance, and cybersecurity. This issue is particularly acute for tabular data, which exhibits heterogeneity, sparsity, nontrivial feature interactions, and lacks spatial structure. This paper presents a systematic survey of intrinsically interpretable deep learning (IIDL) methods tailored for tabular data. We propose a two-dimensional framework that integrates method-centric analysis across pre-model, in-model, and post-model interpretability stages with systems-centric dimensions of timing, target, scope, and audience. Core IIDL architectures, including additive, attention-based, concept-driven, prototype-driven, and hybrid models, are critically reviewed and benchmarked using representative methods such as Neural Additive Models (NAM) and TabNet. We rigorously examine foundational assumptions underlying each architectural family, identifying critical validity constraints including additivity limitations, stability-sparsity trade-offs, supervision requirements, geometric constraints, and computational complexity barriers. Comparative analysis with post-hoc approaches such as SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) highlights their respective robustness and interpretability trade-offs. The survey identifies open challenges in adversarial robustness, causal fidelity evaluation, human-centered interpretability, and mitigation of architectural assumption violations in deployment. Overall, this work provides a methodological and systems-level roadmap for responsible, transparent, and trustworthy AI deployment in regulated environments.