In many machine learning tasks, we don't just want to know how different two distributions are; we want to know how to transform one into the other. This is the essence of the Monge problem: finding a deterministic map $T$ that pushes a source distribution $\rho_a$ to a target $\rho_b$ while minimizing the "effort" (cost) of moving mass.
While Optimal Transport (OT) has long been used as a loss function—most famously in Wasserstein GANs—recent research has shifted toward Neural Optimal Transport (NOT). Instead of using OT to provide a scalar signal for a generator, these methods use neural networks to directly learn the optimal transport map itself. This allows for precise, point-to-point transformations in high-dimensional spaces, enabling applications from image restoration to modeling the evolution of biological cells.
The Minimax Engine of Kantorovich Duality
The primary obstacle to solving the Monge problem is its inherent difficulty: finding a map that satisfies the push-forward constraint $T_{\sharp}\rho_a = \rho_b$ is computationally grueling. To bypass this, researchers leverage Kantorovich Duality. This mathematical framework transforms the search for a map into a saddle-point (minimax) optimization problem.
In this setup, two neural networks compete: a generator (the transport map) and a dual potential (often called a critic or discriminator). As shown by Yujia Xie, Tuo Zhao, and colleagues in their SPOT framework, the dual potential acts as a Lagrange multiplier that enforces the marginal constraints. The generator tries to move samples at minimum cost, while the potential identifies where the moved samples fail to match the target distribution.
This adversarial approach allows OT to scale to high-dimensional data like images. Alexander Korotin and Evgeny Burnaev have demonstrated that these minimax solvers can learn continuous mappings that outperform traditional discrete solvers, which are often limited by the "curse of dimensionality."
Enforcing Optimality with Convexity
A map that moves mass from A to B isn't necessarily optimal. For the specific case of the quadratic Wasserstein-2 distance ($W_2$), Brenier’s Theorem provides a powerful shortcut: the optimal transport map is always the gradient of a convex function.
To bake this mathematical guarantee into the architecture, Ashok Vardhan Makkuva, Amirhossein Taghvaei, and their team introduced the use of Input Convex Neural Networks (ICNNs). By restricting the potential network to be convex, the resulting transport map is automatically regularized toward the unique optimal solution. This structural constraint makes the training process significantly more stable than standard GANs and ensures that the learned map is the most efficient possible transformation.
Handling Mass Non-Conservation
Standard OT assumes that the total "mass" of the two distributions is equal—every point in the source must find a home in the target. In real-world scenarios, like single-cell genomics where cells might divide or die, this assumption fails.
To address this, Caroline Uhler and Karren D. Yang developed a scalable framework for Unbalanced Optimal Transport (UOT). Their method learns both a transport map and a scaling factor simultaneously. This allows the model to "create" or "destroy" mass where necessary, making it robust to outliers and class imbalances that would otherwise distort a standard OT mapping.
From Dynamics to Barycenters
The ability to learn transport maps opens the door to modeling complex systems over time. The JKO (Jordan-Kinderlehrer-Otto) scheme views the evolution of a population as a sequence of optimal transport steps.
- Population Dynamics: Mikhail Persiianov and Alexander Korotin introduced iJKOnet, which uses inverse optimization to recover the hidden "energy landscapes" driving cell development or financial assets from mere snapshots of the population.
- Geometric Averaging: When we have multiple distributions and want to find their "average," we compute a Wasserstein Barycenter. Jiaojiao Fan and Yongxin Chen showed that ICNNs can scale this computation to high-dimensional image data, allowing for the creation of "geometric averages" that preserve sharp features instead of blurring them like a pixel-wise mean would.
The Shift Toward Non-Minimax Solvers
While minimax optimization is powerful, it is notoriously difficult to tune. The latest frontier in this research area involves moving toward non-minimax objectives.
Arip Asadulaev and Alexander Korotin have pioneered methods like U-LightOT and Light Unbalanced OT. By applying clever changes of variables to the dual formulation, they convert the adversarial game into a direct minimization problem. This shift eliminates the instability of competing networks, allowing these solvers to converge faster and more reliably while maintaining the theoretical rigor of the optimal transport framework.