A research team at Tohoku University and Future University Hakodate announced on Apr. 3 that they have successfully trained living biological neurons to carry out a supervised temporal pattern learning task, a function previously performed by artificial systems. The team integrated cultured neuronal networks into a machine learning framework, demonstrating that these biological systems can generate complex time-series signals.
The study, published online in Proceedings of the National Academy of Sciences (PNAS) on March 12, highlights the potential for biological neural networks (BNNs) to serve as alternatives or complements to existing machine learning models. This work marks an important development at the intersection of neuroscience and computational technology.
Artificial neural networks (ANNs) and spiking neural networks (SNNs) are commonly used in machine learning and neuromorphic hardware. Reservoir computing is one approach that leverages the dynamic properties of these recurrently connected systems for processing time-dependent data. In traditional ANN-based reservoir computing, methods like First-Order Reduced and Controlled Error (FORCE) learning allow real-time adaptation by continuously adjusting outputs based on errors, enabling artificial systems to generate various temporal patterns including periodic and chaotic signals.
To explore whether similar approaches could be applied to biological neural networks, researchers constructed BNNs using cultured rat cortical neurons within a reservoir computing framework. By applying FORCE learning to optimize output layers, they trained these biological networks to produce complex temporal signals similar to those involved in motor control. The use of microfluidic devices allowed precise guidance of neuronal growth and control over network connectivity, resulting in modular architectures that minimized excessive synchronization while promoting rich dynamics needed for effective reservoir computing.
The BNN-based system generated diverse time-series patterns such as sine waves, triangular waves, square waves, and chaotic trajectories like the Lorenz attractor. The network also demonstrated flexibility by stably reproducing sine waves with varying periods within the same system.
Looking forward, the research team plans to improve signal stability after training by reducing feedback delays and refining their algorithms. They also aim to expand this platform into a microphysiological system for drug response studies and modeling neurological disorders.
