Brain-inspired sparsity and timing solve AI's catastrophic forgetting

A new preprint from Qianqian Shi and 8 co-authors (arXiv:2605.10178) investigates how the brain flexibly reconfigures neural representations during context transitions without forgetting prior knowledge—a challenge known as catastrophic forgetting in AI. By recording from mouse medial prefrontal cortex (mPFC) and building computational models, the team found that two key mechanisms work together: sparse coding (only a small fraction of neurons are active for a given context) and temporal dynamics (activity patterns evolve over time). Sparsity reduces overlap between context representations, minimizing interference, while temporal dynamics further separate those representations across time, making them more distinct.

When these properties were combined in spiking neural networks, the models achieved significantly better retention during sequential learning tasks—without needing auxiliary tricks like memory replay. The authors argue that this joint mechanism is inherently energy-efficient because it constrains activity, offering a principled way to design lifelong learning systems. The work bridges neuroscience and machine learning, suggesting that building sparsity and temporal structure into artificial networks could lead to more adaptive, stable, and power-efficient AI that learns continuously without overwriting previous knowledge.