Researchers at Tohoku University and Future University Hakodate have trained living rat cortical neurons to perform supervised machine learning tasks in real time, generating complex temporal signals that were previously the exclusive domain of artificial systems. The work, published in Proceedings of the National Academy of Sciences on March 12, represents the first demonstration that cultured biological neural networks can be integrated into a reservoir computing framework and taught to produce specific output patterns on demand.
The team, led by Hideaki Yamamoto, professor at Tohoku University, cultured rat cortical neurons in microfluidic devices that precisely guided neuronal growth and controlled network connectivity. The devices created modular network architectures — a design choice that proved critical. Unstructured neuron cultures tend toward excessive synchronisation, which flattens the high-dimensional dynamics needed for computation. The microfluidic compartmentalisation suppressed that synchronisation while preserving the rich, varied firing patterns the system required.
FORCE learning closes the loop
The neurons were coupled with high-density microelectrode arrays and integrated into a closed-loop system using FORCE (First-Order Reduced and Controlled Error) learning, a training algorithm that adjusts a readout layer’s weights based on the difference between actual and target outputs. In practice, the system listened to the neurons’ spontaneous activity, identified which firing patterns correlated with the desired signal, and iteratively reinforced those correlations.
The approach worked. The biological neural network generated periodic signals — sine waves with periods ranging from 4 to 30 seconds, triangle waves, and square waves — as well as chaotic trajectories matching the Lorenz attractor, a benchmark test for dynamical systems. The ability to produce multiple waveform types within a single neuronal culture suggests the networks contain sufficient computational diversity to support a range of tasks.
Limitations and what comes next
Performance degraded once training ended and the system ran autonomously: mean squared error increased in 99 per cent of trials after the feedback loop was removed. The system also carried a roughly 330-millisecond feedback latency, which limited its ability to track fast-changing or sharp-edged waveforms. These constraints are significant — any practical application would need the system to maintain accuracy without continuous correction and to operate at speeds relevant to real-world control tasks.
The team plans to address both issues by reducing feedback delays through specialised hardware and refining the FORCE algorithm for better post-training stability. Yamamoto noted that the work opens a path toward using living neuronal networks as novel computational resources, with potential applications in brain-machine interfaces and neuroprosthetic devices — though those remain speculative at this stage.
Context: biocomputing meets BCI
The study sits at the intersection of two active fields. Reservoir computing, which treats a complex dynamical system as a fixed “reservoir” and only trains the output layer, has been applied to everything from photonic chips to mechanical oscillators. Meanwhile, the broader organoid intelligence movement — exemplified by Cortical Labs’ DishBrain system, which trained neurons to play Pong in 2022 — has been working to demonstrate that biological substrates can perform meaningful computation.
What distinguishes the Tohoku work is the closed-loop, real-time training of a biological system to produce specific, complex temporal outputs, not just respond to stimuli. If the stability and latency problems can be solved, the approach could eventually inform hybrid biological-electronic systems where living neurons handle the pattern generation that silicon struggles with — precisely the kind of signal production that drives motor neuroprosthetics.
The research was supported by the Japan Science and Technology Agency and the Japan Society for the Promotion of Science.