Researchers have developed a bio-AI system in which living rat cortical neurons are trained to carry out real-time computational tasks. The study explores the use of biological neural networks as functional computing systems by combining them with machine learning techniques through a closed-loop reservoir computing approach.
The system integrates living neurons with high-density microelectrode arrays and microfluidic devices. Neural activity is recorded and converted into continuous outputs, which are then fed back into the system as electrical stimulation.
This feedback loop operates with a delay of approximately 330 milliseconds. A real-time learning method continuously adjusts outputs to match target signals, allowing the system to learn without external intervention.
To improve efficiency, researchers organized neurons into 128 micropores connected by microchannels. This structure reduces excessive synchronization, a common issue in unstructured neural networks.
As a result, neuron correlation dropped from 0.45 to around 0.12, enabling more complex and efficient network behavior. Among the tested designs, the lattice network structure delivered the strongest performance.
The system demonstrated the ability to generate multiple waveform patterns, including sine, square, and triangular waves across different time intervals.
It also showed the capacity to approximate complex chaotic systems such as the Lorenz attractor. During training, the system maintained high accuracy, achieving correlation levels above 0.8.
The system’s performance declines once training stops, with errors increasing during autonomous operation. A key limitation is the 330-millisecond feedback delay, which restricts its ability to process rapidly changing signals.
Future work will focus on reducing latency using specialized hardware. Researchers indicate that this technology could support developments in brain-machine interfaces, neural prosthetics, and bio-hybrid AI systems.
