Reproduce “Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation” on OpenResearch

About this paper

Yushun Tang, Ce Zhang, Heng Xu, Shuoshuo Chen, Jie Cheng, Luziwei Leng, Qinghai Guo, Zhihai He

Neuro-Modulated Hebbian Learning (NHL) offers a novel framework for fully test-time adaptation, integrating feed-forward Hebbian plasticity with a feedback neuromodulator to effectively learn early layer representations and mitigate performance degradation from data distribution shifts. The method achieves state-of-the-art classification error reductions across various datasets like CIFAR-10C, CIFAR-100C, ImageNet-C, and digit recognition tasks.

Key insights

• Combining unsupervised, local, feed-forward Hebbian plasticity with global, task-driven feedback from a neuromodulator creates a robust hybrid learning paradigm for test-time adaptation.
• Adapting early layer representations is crucial for mitigating domain shifts, particularly those caused by common image corruptions or scene changes.
• Framing the adaptation process within the free-energy principle provides a principled way to integrate bottom-up generative modeling (Hebbian) with top-down predictive error minimization (neuromodulator).

Summary by alphaXiv.