Neuroscience breakthrough reveals how the brain optimises its neural coding

A new theoretical framework in neuroscience demonstrates that many neural coding problems can be understood through convex optimisation. The research, published recently, bridges mathematical theory with biological realism in a way that avoids the pitfalls of oversimplification or excessive complexity. Scientists believe this approach will accelerate progress in both neuroscience and artificial intelligence.

The study introduces a method that shifts focus from analysing raw neural activity to optimising representational similarity. By doing so, the team uncovered previously unknown links between different coding schemes in the brain. Their findings show that, under specific conditions, the tuning of individual neurons aligns perfectly with the most efficient representational structure.

One key insight explains why the retina uses dense coding while the cortex relies on sparse coding. The research reveals that both strategies optimally divide the processing of a single variable into ON and OFF channels. This split leverages a tractable nonlinearity found in convex problems, offering a clearer picture of how visual information is organised at different stages.

The framework also provides the first complete set of conditions for solving semi-nonnegative matrix factorisation—a tool vital for studying modularity and mixed-selectivity in the entorhinal cortex. Additionally, the team proved that a wide range of optimisation problems, including those in linear and nonlinear neural networks, can be expressed as convex functions over representational similarity matrices.

This work equips researchers with a powerful new toolkit for understanding neural coding. By balancing mathematical rigour with biological plausibility, the framework moves beyond descriptive analyses toward predictive theories grounded in optimality. The findings are expected to influence both neuroscience and AI, offering fresh ways to interpret how the brain processes information.