Study Reveals How Low-Dimensional Neural Codes Enable Reliable Working Memory Despite Neuronal Noise

The first study, posted to arXiv, proposes a quasi-reverse-martingale theory to formalize when an RNN's hidden state is probabilistically stable, introducing 'backward coherence' as a measure of how well a hidden state can be reconstructed from its successor. The authors show that a backward-coherence regularization loss reduces empirical instability by 43–58% and achieves stable representations 28–44% earlier than unregularized baselines, with the loss mathematically equivalent to minimizing a KL divergence in a Gaussian backward model. Validation on three real-world datasets — ICU mortality prediction, macroeconomic forecasting, and human activity recognition — demonstrates practical benefits including fourfold reduction in forecast error under concept drift. The second study, posted to bioRxiv, addresses biological working memory in noisy neural populations, proving analytically that low-dimensional latent manifolds suppress independent neuronal noise while inducing correlated noise that limits downstream information extraction. Their key result is an analytical bound showing working memory duration scales linearly with network size, a prediction they validate with large-scale neocortical recordings and a behavioral signature in mice. While the two papers use different formalisms and study different systems, both converge on the idea that low-dimensional or structured internal representations are critical for sustaining reliable computation over time in recurrent systems.