Summary: The boundary between biology and computer science has moved closer: researchers have trained living rat neurons to perform complex machine-learning tasks. By embedding cultured neuronal networks in a reservoir computing framework and applying an online training method, the team demonstrated that living neural tissue can serve as a real-time computational resource.
Using an approach known as FORCE learning, the researchers taught biological circuits to produce intricate temporal patterns — including the chaotic Lorenz attractor — showing that living “wetware” can generate and maintain complex time-series signals similar to those produced by artificial systems.
Key facts
- Reservoir computing: This strategy exploits the inherent, high-dimensional dynamics of a recurrent network (the “reservoir”) to process time-dependent data. Instead of training every neuron, only the readout layer that interprets the network’s activity is trained.
- FORCE learning: First-Order Reduced and Controlled Error (FORCE) learning adapts outputs in real time by minimizing error. This study is the first to apply FORCE learning successfully to a biological neural network (BNN) to generate temporal patterns.
- Chaotic dynamics: The living networks reproduced not only simple periodic signals but also complex chaotic trajectories, notably the Lorenz attractor, a benchmark for chaotic time-series often used to model weather and other nonlinear systems.
- Microfluidic control: Microfluidic devices guided neuronal growth to form modular “neighborhoods” of cells. This prevented excessive synchronization and promoted the diverse, high-dimensional activity needed for effective reservoir computing.
- Functional versatility: The same biological system learned and stably reproduced waves with periods from 4 to 30 seconds, demonstrating adaptability across multiple time scales.
Source: Tohoku University
Overview
A research team from Tohoku University and Future University Hakodate showed that cultured rat cortical neurons can be trained to perform supervised temporal pattern learning tasks previously demonstrated only with artificial networks. By combining living neuronal cultures, microfluidic patterning, high-density microelectrode arrays, and an online feedback-learning algorithm, the investigators created a closed-loop biological reservoir computing system that generates coherent temporal outputs.
The study, published in Proceedings of the National Academy of Sciences (PNAS) on March 12, 2026, presents strong evidence that biological neural networks (BNNs) can be used as computation-capable substrates. These results bridge experimental neuroscience and bio-inspired computing, suggesting that living networks might complement or offer alternatives to traditional machine-learning models for time-dependent tasks.
Reservoir computing has been an effective approach for processing temporal data with artificial neural networks (ANNs) and spiking neural networks (SNNs). In conventional artificial systems, FORCE learning and related adaptive methods enable a readout layer to learn and produce periodic and chaotic signals in real time. Until now, applying such strategies directly to living neural tissue remained an open question.
To test this, the researchers grew cortical neurons from rats in vitro and arranged them within microfluidic chambers that guided connectivity and limited global synchronization. Integrating these cultures with high-density microelectrode arrays allowed real-time monitoring and closed-loop feedback. Training used a linear decoder with fixed feedback weights, adjusted online by the FORCE algorithm to shape output signals from the BNN.
When feedback was applied, the originally irregular neuronal activity converged to low-dimensional, structured dynamics, producing coherent trajectories with stable transitions between neural states. The trained BNNs reproduced a variety of temporal signals — sine, triangular, and square waves — and, importantly, chaotic trajectories like the Lorenz attractor. The system’s ability to learn oscillations with periods between 4 and 30 seconds demonstrates its flexibility across multiple temporal regimes.
Hideaki Yamamoto, a professor at Tohoku University, emphasized the broader implications: “This work shows that living neuronal networks are not only biologically meaningful systems but may also serve as novel computational resources. By bridging neuroscience and machine learning, we are opening a pathway toward new forms of computing that leverage the intrinsic dynamics of biological systems.”
Next steps for the team include improving the stability and persistence of generated signals after training, reducing feedback delays, and refining the FORCE algorithm for biological contexts. The platform could also be adapted as a microphysiological system to study drug effects and model neurological disorders, expanding its value for both computing research and biomedical applications.
Key questions answered
A: These experiments belong to a field sometimes called “wetware computing.” Unlike silicon chips, biological reservoirs exploit the noisy, energy-efficient dynamics of living cells to process information. They offer adaptability and parallelism that differ from conventional AI, though they are not replacements for silicon systems today.
Q: How do researchers “teach” a dish of cells to compute?
A: Think of the network as an orchestra already producing diverse activity. FORCE learning acts like a conductor and a feedback system that reinforces the combinations of activity that produce a desired output. Over time, the readout learns which network signals to use to generate the target pattern.
Q: Why use real neurons instead of standard AI?
A: Biological networks excel at massively parallel processing with low energy consumption and can naturally produce rich dynamics suitable for temporal tasks. They also offer experimental platforms to test pharmacological effects on network computation and to model disease-related circuit dysfunctions.
Editorial notes
- This article was edited by a Neuroscience News editor.
- The journal paper was reviewed in full.
- Additional context was provided by staff.
About this AI and neuroscience research news
Author: Public Relations Office
Source: Tohoku University
Contact: Public Relations Office – Tohoku University
Image: The image is credited to Neuroscience News
Original research: Open access. “Online supervised learning of temporal patterns in biological neural networks under feedback control” by Yuki Sono, Hideaki Yamamoto, Yusei Nishi, Takuma Sumi, Yuya Sato, Ayumi Hirano-Iwata, Yuichi Katori, and Shigeo Sato. PNAS. DOI: 10.1073/pnas.2521560123
Abstract
Online supervised learning of temporal patterns in biological neural networks under feedback control
In vitro biological neural networks (BNNs) provide controlled model systems to study how living cells interact with their environment to create high-dimensional dynamics that can be harnessed to generate coherent temporal outputs relevant for behaviors such as motor control. Here, the authors develop a real-time closed-loop BNN system that integrates cultured cortical neurons with microfluidic devices and high-density microelectrode arrays to produce periodic and chaotic temporal signals.
Training a simple linear decoder with fixed feedback weights enables the system to learn and autonomously generate diverse temporal patterns. When feedback is engaged, irregular BNN activity transforms into low-dimensional, structured dynamics with coherent trajectories and stable transitions between neural states. BNNs trained on target frequencies ranging from 4 to 30 seconds can sustain oscillations at distinct frequencies, showing adaptability across time scales. Controlling self-organized network formation via microfluidic patterning suppresses excessive synchronization and increases dynamic complexity, which facilitates training and yields robust outputs. This platform advances understanding of cortical computations and supports development of energy-efficient neuromorphic paradigms inspired by biology.