Communications Biology · 2026 · Journal article

Representation Transfer via Invariant Input-driven Continuous Attractors for Fast Domain Adaptation

Tie Xu, Shengdun Wu, Junwen Luo, Yao Yang, Feng Lin, Junjie Jiang, Guozhang Chen, Dongping Yang

Modular recurrent networks form invariant input-driven continuous attractors that enable rapid, few-shot domain adaptation through lightweight readout adapters.

The scientific question

How can neural representations transfer reliably across shifting, noisy real-world domains?

End-to-end deep networks degrade under environmental shifts and require expensive full-model retraining when deployed in noisy, changing conditions.

Concept, pretraining and emergent neural manifolds for representation transfer.
Figure 1 from the original paper ↗ · CC BY-NC-ND 4.0

The core idea

Modular Continuous Attractor Pretraining

Task-agnostic pretraining shapes modular recurrent networks to embed inputs onto low-dimensional, invariant continuous attractor manifolds, requiring only a lightweight adapter for new downstream tasks.

What we found

Parameter-Efficient Few-Shot Adaptation

On gesture and rehabilitation action recognition tasks, the framework achieves accuracy competitive with state-of-the-art methods, especially in few-shot settings, while requiring an order of magnitude fewer parameters and minimal training.

Continue the question

Related reading

Explore our research

Cite this work

Tie Xu, Shengdun Wu, Junwen Luo, Yao Yang, Feng Lin, Junjie Jiang, Guozhang Chen, Dongping Yang. Representation Transfer via Invariant Input-driven Continuous Attractors for Fast Domain Adaptation. Communications Biology, 2026. doi:10.1038/s42003-026-09938-8.