monotone disjunctive-normal-form (MDNF)

Gold definitionUpdated Apr 2, 2026

Definition

Monotone Disjunctive-Normal-Form (MDNF) is a type of logical formula used to generate interpretable explanations for black-box AI models. Its satisfaction guarantees a high model score for a given class, expressing explanations in terms of human-recognizable primitive concepts.

At a glance

Executive summary

Monotone Disjunctive-Normal-Form (MDNF) is a method for creating understandable explanations for complex AI models, especially deep neural networks. It translates a model's decisions into simple logical rules based on human-friendly concepts, showing why a model made a specific prediction with high accuracy.

TL;DR

MDNF helps explain why complex AI models make certain decisions by turning their internal logic into easy-to-understand rules using simple concepts.

Key points

Casts model explanations as logical formulas (OR of ANDs of primitive concepts) whose satisfaction guarantees a high model score for a class.
Addresses the opacity of black-box AI models, providing human-interpretable insights into their decision-making processes.
Used by researchers and ML engineers in explainable AI, particularly for computer vision tasks and understanding deep neural networks.
Unlike feature importance methods that assign scores, MDNF provides structured, rule-based explanations that directly map to logical conditions.
Represents a growing interest in generating inherently interpretable models or post-hoc explanations that maintain high fidelity and coverage for complex AI systems.

Use cases

Medical Diagnosis: Explaining why an AI model diagnosed a patient with a certain condition based on a combination of symptoms and medical image features.
Autonomous Driving: Providing logical reasons for a self-driving car's decision, such as 'IF pedestrian detected AND traffic light is red THEN brake,' enhancing trust and debugging.
Fraud Detection: Uncovering the specific logical rules (e.g., 'IF transaction amount > X AND location is Y AND time is Z THEN flag as fraud') that an AI uses to identify suspicious activities.
Image Classification Debugging: Understanding why a deep learning model misclassified an image by seeing the MDNF explanation for the incorrect class, revealing biases or errors in learned features.

Also known as

MDNF, monotone explanation list