Learning to maintain information over long periods is a fundamental challenge for both brains and artificial neural networks. A new paper from researchers at Harvard University and Stanford University, published on arXiv in March 2025, builds a mathematical theory of how linear recurrent neural networks (RNNs) learn to integrate white noise — a task that requires holding onto input over time.
What the Researchers Did
Blake Bordelon (Harvard), Jordan Cotler (Harvard), Cengiz Pehlevan (Harvard), and Jacob A. Zavatone-Veth (Harvard and Stanford) trained linear RNNs to integrate white noise, a signal where each time step is independent and random. They started with small recurrent weights and tracked how the network's connectivity evolved during gradient-based learning — the same algorithm used in training modern AI.
They discovered that the learning dynamics collapse to a low-dimensional system tracking a single "outlier" eigenvalue of the recurrent weight matrix. This eigenvalue grows over training, creating a long intrinsic timescale that allows the network to integrate inputs over many time steps. The process precisely mirrors how a continuous attractor or integrator model maintains persistent activity.
In a second set of experiments, the team extended the analysis to networks learning a damped oscillatory filter. Here, a conjugate pair of eigenvalues evolves, producing oscillatory dynamics with a specific frequency and decay rate.
Why It Matters for Your Brain
This mathematical framework shows a clear path from initial random connectivity to a highly tuned integrator — a circuit that can hold onto information for seconds or longer. While the paper studies artificial neurons, the principles likely apply to biological neural circuits, such as those in the prefrontal cortex that support working memory. Understanding these dynamics could help cognitive scientists design better protocols for memory training and help individuals optimize their own cognitive performance.
Actionable Takeaway
While you can't directly tweak your brain's eigenvalues, you can strengthen your working memory through practice. Tasks that require holding and manipulating information — like mental arithmetic, memorizing sequences, or playing n-back games — are known to engage these integrator circuits. Even simple dual-task training, such as walking while solving math problems, can enhance your brain's ability to maintain and integrate information over time.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
