Unbalancedness in Neural Monge Maps Improves Unpaired Domain Translation
Luca Eyring*, Dominik Klein*, Théo Uscidda*, Giovanni Palla, Niki Kilbertus, Zeynep Akata, Fabian Theis
International Conference on Learning Representations, ICLR


In optimal transport (OT), a Monge map is known as a mapping that transports a source distribution to a target distribution in the most cost-efficient way. Recently, multiple neural estimators for Monge maps have been developed and applied in diverse unpaired domain translation tasks, e.g. in single-cell biology and computer vision. However, the classic OT framework enforces mass conservation, which makes it prone to outliers and limits its applicability in real-world scenarios. The latter can be particularly harmful in OT domain translation tasks, where the relative position of a sample within a distribution is explicitly taken into account. While unbalanced OT tackles this challenge in the discrete setting, its integration into neural Monge map estimators has received limited attention. We propose a theoretically grounded method to incorporate unbalancedness into any Monge map estimator. We improve existing estimators to model cell trajectories over time and to predict cellular responses to perturbations. Moreover, our approach seamlessly integrates with the OT flow matching (OT-FM) framework. While we show that OT-FM performs competitively in image translation, we further improve performance by incorporating unbalancedness (UOT-FM), which better preserves relevant features. We hence establish UOT-FM as a principled method for unpaired image translation.

Neural Monge Maps:

Unpaired domain translation aims to transform data from a source to a target distribution without access to paired training samples. This setting poses the significant challenge of achieving a meaningful translation between distributions while retaining relevant input features. Although there are many ways to define the desired properties of such a transformation, optimal transport (OT) offers a natural framework by matching samples across distributions in the most cost-efficient way.

If this optimal correspondence can be formulated as a map, such a map is known as a Monge map. Recently, a considerable number of neural parameterizations to estimate Monge maps have been proposed.

Incorporating Unbalancedness:

Neural Monge maps have been successfully applied to a variety of domain translation tasks. However, optimal transport assumes a static marginal distribution, which can limit its application as it cannot account for [i] outliers and [ii] undesired distribution shifts, e.g. class imbalance between source and target distribtuion as in the example above. Unbalanced OT (UOT) overcomes these limitations by replacing the conservation of mass constraint with a penalization on mass deviations as seen in the following:


However, existing methods for estimating neural Monge maps with unbalancedness are limited. In light of these limitations, we introduce a new framework for incorporating unbalancedness into any neural Monge map estimator based on a re-scaling scheme. We motivate our approach theoretically and propose the concept of an unbalanced Monge map.


We show that with our method we can mimic the behaviour of discrete UOT for any neural Monge map estimator. This also seamlessly integrates with the OT Flow Matching (OT-FM) framework as shown above.



We demonstrate the importance of learning unbalanced Monge maps on three different domain translation tasks leveraging three different (balanced) Monge map estimators, which showcases the flexibility of our proposed method. We find that unbalancedness enhances the performance in all three of these settings.

In particular, We apply our method to existing estimators across two unpaired single-cell translation settings. Here, we show how unbalancedness is crucial in recovering meaningful single-cell trajectories over time with OT-ICNN and in predicting cellular responses to cell perturbations with Monge Gap. Moreover, we demonstrate that while OT-FM performs competitively in unpaired image translation, unbalancedness (UOT-FM) further elevates these results and better preserves relevant input features.


Lastly, we show that unbalancedness also improves performance in the image generation setting. In particular, a small amount of unbalancedness with UOT-FM improves upon OT-FM on CIFAR-10 image generation. For all the details, check out the paper!

(c) 2023 Explainable Machine Learning Munich Impressum