Active inference framework models decision-making in neural networks mimicking biological brains

Researchers have developed a mathematical framework grounded in active inference—a theory of how agents learn and make decisions—to understand adaptive behavior in systems ranging from artificial neural networks to biological brain tissue in lab settings.

The Challenge: Building Safer, More Efficient AI#

As AI systems become more powerful, understanding the fundamental principles driving adaptation in autonomous agents has become crucial. This knowledge helps researchers build systems that are both effective and safe. While deep learning has traditionally dominated the field, there's growing interest in alternative approaches: living biological neural networks, which appear to offer unusual advantages like exceptional learning efficiency and complex, dynamic behavior patterns that might inspire more transparent and biologically grounded AI models.

A Computational Bridge Between Artificial and Living Systems#

The team constructed an active inference model calibrated to match an experimental protocol using biological neural cultures—specifically, the DishBrain platform, which explores how isolated cortical tissue from mice and humans can learn to control a game-like environment. Using this common mathematical framework, the researchers systematically tested three different decision-making strategies within identically matched generative models.

The key finding emerged when comparing learning under realistic constraints: agents equipped with shorter memory spans matched the performance levels achieved by the biological neural cultures, while those operating with extended memory horizons performed substantially worse. Interestingly, expanding the planning horizon—how far into the future an agent reasons—provided no equivalent performance advantage.

What This Reveals About Learning#

Because all model parameters remain explicit and trackable within the active inference framework, the researchers could directly observe what quantities in the mathematical system changed as performance improved. This included shifts in the risk term (reflecting uncertainty) and variations in the entropy of transition and state-action mappings—essentially, how much randomness or structure existed in the system's learned behaviors.

These findings demonstrate that active inference provides a formal language for comparing fundamentally different decision-making strategies within similar closed-loop control scenarios, bridging computational theory with experimental biology.