Brain cortex study reveals hidden trick to supercharge neural networks

A team of researchers from Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) conducted a simulation-based reverse engineering study on cortical microcircuits to understand how biological neural networks optimize information flux — the mutual information between successive network states, considered critical for rich information processing. They modeled a simplified layer 5 cortical architecture featuring a densely connected core population embedded within a larger, sparsely connected supporting network. Surprisingly, they discovered that the embedding network exerts a pronounced flux-enhancing effect on the core, primarily through two mechanisms: generating effective biases that push core neurons into higher-entropy operating regimes, and supplying stochastic fluctuations that prevent the network from getting trapped in simple fixed-point or oscillatory attractors. This latter phenomenon, which they term "Recurrence Resonance," is a novel dynamical mechanism.

The team further demonstrated that the biological configuration is not the global optimum. By applying individually optimized biases to core neurons, they achieved information flux even higher than the biologically embedded case. These optimal biases can emerge from a simple self-organization principle, suggesting a plausible evolutionary path. The findings have dual relevance: they provide functional insights into how cortical microcolumns process information, and they offer concrete design principles for artificial recurrent systems like reservoir computers. For AI practitioners, this suggests that introducing structured noise and bias modulation into recurrent neural networks could significantly boost computational capacity without increasing model size.