Researchers at Tohoku University and Future University Hakodate have shown that living biological neurons can be trained to perform a supervised temporal pattern learning task that was previously done only by artificial systems. The work integrates cultured neuronal networks into a machine learning setup and demonstrates that these living systems can produce complex time-based signals.
The study is published in PNAS. It suggests that biological neural networks could one day serve as alternatives or additions to current machine learning models.
Biological neurons show computing potential
Spiking neural networks closely mimic how real neurons send quick electrical pulses called spikes. Reservoir computing is an efficient computing method that uses the natural changing states of a network to handle data that varies over time, such as sound or movement signals. FORCE learning is a training technique that adjusts output connections in real time based on errors to help the system match desired patterns.
In the experiment, scientists grew networks of rat cortical neurons, which are nerve cells from the brain's outer layer, inside microfluidic devices. These devices guide cell growth and control how the neurons connect to one another. The setup used microelectrode arrays to record and stimulate the cells. By applying FORCE learning to adjust the readout layer, which turns the network's activity into output signals, the living networks learned to generate various time-series patterns. These included sine waves, triangular waves, square waves, and even chaotic signals such as the Lorenz attractor, a mathematical model of complex, unpredictable motion.
The same network could learn and stably reproduce sine waves with different periods ranging from 4 to 30 seconds. The modular design helped avoid excessive synchronization among neurons and supported the rich dynamics needed for effective computing.
This platform shows that living neuronal networks can act as computational resources that use their own natural activity. Future work may focus on making signal generation more stable after training ends and on reducing delays in feedback.