Online supervised learning of temporal patterns in biological neural networks under feedback control

Abstract

In vitro biological neural networks (BNNs) provide a well-defined model system to constructively investigate how living cells interact with their environment to shape high-dimensional dynamics that could be used to generate a coherent temporal output, such as those required for motor control. Here, we developed a real-time closed-loop BNN system capable of generating periodic and chaotic temporal signals by integrating cultured cortical neurons with microfluidic devices and high-density microelectrode arrays. We show that training a simple linear decoder with fixed feedback weights enables the system to learn and autonomously generate diverse temporal patterns. When feedback was switched on, irregular activity in BNNs is transformed into low-dimensional, structured dynamics, producing coherent trajectories characterized by stable transitions between neural states. BNNs trained on different target frequencies--ranging from 4 to 30 s--could be trained to sustain oscillations at distinct frequencies, demonstrating their adaptability. Importantly, a top-down control of self-organized network formation with microfluidic devices is the key to suppress excessive synchronization and increase dynamical complexity in BNNs, facilitating the training and robust output generation. This work offers a biologically inspired platform for understanding the physical basis of cortical computation and for advancing energy-efficient neuromorphic computation. Significance StatementReservoir computing is a machine learning paradigm that exploits the transient dynamics of high-dimensional nonlinear systems. Although it was originally inspired by the mammalian brain and widely explored in physical systems, its implementations in biological neural networks (BNNs) have been limited due to their excessive connectivity and global synchrony in vitro. Here, we use microfluidic devices to construct modular, nonrandomly connected BNNs and integrate them with microelectrode arrays in a closed-loop reservoir computing environment. We show that the system can be trained to autonomously output various temporal signals, with the modular connectivity that is essential for learning. In vitro BNNs provide unique alternatives for physical reservoirs with dynamic adaptability.