SIDDA: SInkhorn Dynamic Domain Adaptation for Image Classification with Equivariant Neural Networks

The Big Picture

Imagine training a doctor to diagnose X-rays from one hospital, then discovering that scans from another hospital look slightly different: different scanner models, different contrast, different noise. The expertise doesn’t disappear, but accuracy plummets.

Scale that problem to astronomy. A model trained on simulated galaxy images from IllustrisTNG must classify real galaxies from the Hubble Space Telescope. Simulations are clean. Real data is blurred, noisy, and shot through with instrumental artifacts. The neural network, excellent in the lab, stumbles in the field.

This is the training-deployment gap, one of the most persistent headaches in applied machine learning. The underlying task hasn’t changed, and the labels still mean the same thing. But the statistical character of the data has shifted, and modern neural networks are surprisingly brittle when that happens.

Existing fixes, grouped under domain adaptation, generally work but come with a catch: they demand extensive tuning of dozens of internal settings, burning enormous compute and expert time.

A team from Northeastern University, Fermilab, the University of Chicago, and the University of Toronto has found a better approach. Their method, SIDDA (SInkhorn Dynamic Domain Adaptation), achieves effective domain alignment with minimal tuning. Paired with symmetry-aware neural networks, it pushes classification accuracy up by as much as 40% on unlabeled real-world data.

Key Insight: SIDDA automatically aligns training and deployment data distributions using the Sinkhorn divergence, a computationally efficient form of optimal transport. It requires almost no hyperparameter tuning while producing large gains in both accuracy and model calibration.

How It Works

The mathematical heart of SIDDA is optimal transport (OT) theory, originally developed to answer a deceptively simple question: what is the cheapest way to move a pile of dirt from one place to another? In machine learning, the “dirt” is your data’s statistical spread, and the goal is to move it from the training domain to the real-world domain at minimum cost. OT captures not just how different two distributions are, but the geometry of those differences.

Raw OT is expensive. SIDDA uses the Sinkhorn divergence, an efficient approximation that runs orders of magnitude faster while preserving the desirable geometric properties.

What sets SIDDA apart is making this divergence dynamic. Rather than computing alignment once at the start of training, SIDDA continuously updates the domain alignment loss throughout. The network learns domain-invariant features progressively instead of solving a rigid optimization problem upfront.

The training loop works as follows:

1. Feed a batch of labeled source images and unlabeled target images through the network in parallel.
2. Compute the standard classification loss on the source data.
3. Compute the Sinkhorn divergence between the latent representations (the network’s compressed, abstract encoding of each image) for source and target data.
4. Sum both losses with a weighting factor λ and backpropagate.
5. Repeat. The domain gap shrinks progressively as training proceeds.

The λ weight is the one adjustable setting that matters, and it proves stable across datasets. Compare this to Maximum Mean Discrepancy (another distributional distance measure) or adversarial domain adaptation (which pits two networks against each other to force alignment), both of which need careful per-dataset tuning.

SIDDA also pairs well with equivariant neural networks (ENNs), which mathematically enforce symmetry rather than learning it approximately from data. Standard convolutional networks pick up approximate symmetries through training; ENNs bake them in. For galaxy images and other rotationally symmetric objects, this makes a real difference.

The authors test ENNs with dihedral group symmetry $D_N$ (rotational and reflection symmetry at $N$ discrete angles) and find a clean result: the higher the degree of equivariance, the better SIDDA performs. Symmetry and domain adaptation reinforce each other.

Why It Matters

The immediate application is astronomy. Next-generation surveys from the Vera Rubin Observatory’s LSST, the Nancy Grace Roman Space Telescope, and the Euclid mission will produce data volumes no human team can manually classify. Pipelines trained on simulations will be deployed on real sky, and every percentage point lost to domain shift is a percentage point of scientific insight lost.

SIDDA closes that gap without requiring each new survey team to spend months tuning adaptation parameters.

The method also works outside astronomy. The paper tests SIDDA on simple shapes, handwritten digits, and real astronomical observations, a deliberately varied set of problems. Medical imaging faces similar challenges when MRI machines from different manufacturers produce subtly different scans. Autonomous driving runs into the same issue when weather changes. Anywhere a model is trained in one environment and deployed in another, this kind of adaptation matters.

Open questions remain. The current framework is unsupervised on the target side, assuming no labeled target data at all. Semi-supervised extensions, where even a handful of target labels are available, could push accuracy higher. And while the paper studies dihedral symmetry groups, the interaction between continuous symmetry groups and Sinkhorn alignment is still largely uncharted.

Bottom Line: SIDDA achieves up to ~40% classification accuracy gains and over an order of magnitude improvement in calibration error on unlabeled target data, with almost no hyperparameter tuning, making reliable neural networks practical for astronomy and beyond.