How does your brain hold a phone number in mind for a few seconds? New research points to a specific neural computation that helps maintain such continuous working memory.
The Research
A team led by Zhaotian Gu at Fudan University and collaborators from multiple institutions (submitted to arXiv on August 3, 2026) studied how artificial recurrent neural networks (RNNs) can learn to maintain continuous variables—a hallmark of working memory. Traditional continuous attractor networks are fragile and require fine-tuning, while popular RNNs like GRUs and LSTMs often fail, breaking the continuous space into discrete points.
Inspired by biology, the researchers proposed the Recurrent Divisive Normalization Network (RDNN), which incorporates divisive normalization—a computation seen in cortical circuits. Through dynamical systems analysis on canonical working memory tasks, they found that RDNNs converge to robust, high-fidelity slow manifolds, meaning they can smoothly maintain memory values. They also analyzed the gradient dynamics during training, showing that divisive normalization leads to activity-dependent gradient scaling that compresses the network's effective rank, confining dynamics to a low-dimensional subspace without explicit rank constraints.
Importantly, ablation experiments revealed that while subtractive inhibition can hold static memories, divisive normalization is mathematically essential to prevent 'manifold shattering' when inputs change over time. This suggests that divisive normalization is not just a biological curiosity but a key mechanism for learning continuous representations.
Why It Matters
For your own brain, this research highlights the importance of normalizing neural activity—something your brain does all the time—to keep information stable. It may explain why some people are better at maintaining focus or recalling details under distraction. For AI, it suggests better architectures for tasks requiring continuous memory, such as navigation or language understanding.
What You Can Do
While you can't directly manipulate your brain's divisive normalization, you can train your working memory with exercises that require maintaining and updating information, such as mental math, remembering sequences, or dual-task training. Practice focusing on a single task for extended periods without interruption.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
