Game theory reveals stability in asymmetric excitatory-inhibitory neural circuits
Neurons act as self-interested agents to maintain balance—a new framework for biologically realistic AI.
Energy-based models have been a cornerstone for understanding computation and stability in neural systems, but they traditionally require symmetric weight matrices—a condition absent in real biological E-I networks. This asymmetry breaks the concept of a global energy landscape, leaving the dynamics of such networks poorly understood. In a new arXiv preprint, Simone Betteti and colleagues from UC San Diego propose a game-theoretic extension: each neuron acts as an agent minimizing its own local energy. This reinterpretation preserves rigorous stability analysis via network theory, enabling the study of regulation and balancing in E-I circuits without artificial symmetry.
Applying their framework to Wilson-Cowan and lateral inhibition models, the team shows how cortical columns function as contrast enhancers. By modeling hierarchical excitation-inhibition interactions, they demonstrate that subtle differences in input can be selectively amplified—a key capability for sensory processing. The results bridge energetic and game-theoretic views, providing a systematic method to design dynamically stable neural networks that are biologically grounded. This opens the door to more robust and interpretable AI architectures inspired by the brain's own balancing acts.
- Removes symmetry constraint of traditional energy-based models by introducing game-theoretic neuron objectives.
- Each neuron minimizes its own energy, enabling stability analysis for asymmetric excitatory-inhibitory networks.
- Applied to lateral inhibition microcircuits, showing how hierarchical E-I interplay sharpens sensory contrast.
Why It Matters
Bridges neuroscience and game theory, offering a blueprint for designing stable, biologically plausible neural networks for AI.