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

Conventional end-to-end deep neural networks often degrade under domain shifts and require costly retraining when deployed in unpredictable, noisy environments. Inspired by biological brains, we propose a modular framework where each module is a recurrent neural network pretrained via a simple, task-agnostic protocol to learn robust, transferable features. This shapes stable yet flexible low-dimensional representations as invariant input-driven continuous attractor manifolds embedded in high-dimensional latent space across different tasks, supporting robust transfer and resilience to temporal perturbations. At deployment, only a lightweight adapter needs training, allowing rapid adaptation to new tasks. Validated on gesture and rehabilitation action recognition tasks, our 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. By integrating biologically inspired attractor dynamics with cortical-like modular composition, the framework offers a practical path toward robust, continual adaptation in real-world information processing.