Brain-inspired Representation Transfer through Invariant Input-driven Continuous Attractors in a Modular RNN Framework
Its authors withdrew this paper, so there is no full text to read or summarize.
Lists
Only you see this
Abstract
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 neural computation, we propose a modular framework in which each module is a recurrent neural network pretrained using a simple, task-agnostic protocol to learn robust, transferable features. We show that low-dimensional, input-driven continuous attractor manifolds, embedded in a high-dimensional latent space, yield task-invariant representations that enable robust transfer and resilience to temporal perturbations. At deployment, only a lightweight adapter needs training, allowing rapid adaptation to new tasks. Validated on the Dynamic Vision Sensor (DVS) Gesture benchmark and a custom rehabilitation action recognition dataset we collected, 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 provides a practical route to robust, continual adaptation in real-world information processing.
BibTeX
@misc{xu2026brain,
author = {Tie Xu and Shengdun Wu and Junwen Luo and Feng Lin and Dongping Yang},
title = {{Brain-inspired Representation Transfer through Invariant Input-driven Continuous Attractors in a Modular RNN Framework}},
year = {2026},
url = {https://openreview.net/forum?id=XZZ4En7yOK},
howpublished = {OpenReview},
note = {Submitted to ICLR 2026},
}Loading related content...