- Biological substrate
- 2D modular cortical culture
- Cells
- Cultured cortical neurons
- Interface
- Microfluidics + high-density MEA
- Task
- Autonomous temporal-pattern generation
- Evidence
- Peer reviewed
- DOI
- 10.1073/pnas.2521560123
Sono and colleagues built a real-time closed-loop biological neural network from cultured cortical neurons, microfluidic connectivity structures and a high-density microelectrode array. A linear decoder was trained while fixed feedback weights returned the decoded output to the culture.
The system learned to generate periodic signals at several target frequencies and also produced a Lorenz-attractor trajectory. Under suitable conditions, the trained network continued generating the target dynamics after the learning phase ended.
Why the microfluidics matter
Dense dissociated cultures can become excessively synchronized, which reduces the variety of useful internal dynamics. The study used microfluidic structures to impose modular, non-random connectivity. The authors report that this network organisation increased dynamic complexity and was important for training.
What the experiment establishes
This is a peer-reviewed demonstration of online supervised learning in an in-vitro BNN coupled to a digital feedback loop. The trained digital decoder remains part of the computational system, so the result should be read as a hybrid biological–electronic architecture.
What to measure next
Cross-culture repeatability, longer retention tests, matched electronic reservoir baselines and complete system-energy measurements would make the engineering significance easier to judge.