MDD

Gold definitionUpdated Apr 2, 2026

Definition

MDD primarily refers to Margin Disparity Discrepancy, an unsupervised domain adaptation (UDA) framework that bridges modality gaps by reformulating optimization. It also denotes a specific type of Multi-valued Decision Diagram (MDD) used in constraint programming for no-overlap scheduling.

At a glance

Executive summary

MDD is a technique used in AI to adapt models trained on one type of data to work well on a different but related type of data, especially when the new data lacks labels. Separately, it refers to a tool in computer science for efficiently solving complex scheduling problems by ensuring tasks don't overlap.

TL;DR

MDD helps AI models learn from different kinds of data without needing new labels, and also helps solve complex scheduling puzzles by figuring out when tasks can happen without clashing.

Key points

Margin Disparity Discrepancy (MDD) is a formalism for unsupervised domain adaptation, reformulating optimization to bridge modality gaps.
It addresses data scarcity and lack of annotation in target domains, especially in medical imaging, and provides bound-consistent filtering for no-overlap constraints in scheduling.
Used by researchers in medical imaging, machine learning engineers for domain adaptation, and constraint programming specialists for scheduling.
In UDA, it offers state-of-the-art performance compared to other domain adaptation techniques. In scheduling, it provides a bound-consistent algorithm for no-overlap, outperforming previous precedence-detection methods.
Growing importance in leveraging unlabeled data for robust AI models across domains (UDA) and developing efficient constraint satisfaction algorithms for complex real-world problems.

Use cases

Liver segmentation in interventional Cone-Beam CT (CBCT) using models trained on annotated CT data.
Adapting diagnostic AI models from standard, labeled datasets to proprietary, unannotated clinical data in healthcare.
Optimizing job sequencing with time windows to minimize tardiness and earliness in manufacturing or logistics.
Resource allocation problems where tasks cannot overlap, such as scheduling operating rooms or production lines.
Developing robust computer vision systems that perform well across varying imaging conditions or sensor types without extensive re-labeling.

Also known as

Margin Disparity Discrepancy, no-overlap MDD, relaxed MDD

MDD