Understanding neural communication and function requires efficient analysis of biological neural network activities. The firing patterns and collective activities of neurons govern functions like consciousness and memory formation. Current neural probe technologies (patch clamp, nanowire-based FETs, MEMS probes, CMOS nanoelectrode arrays) record intracellular and extracellular electrophysiological activities (neural spikes). The massive amounts of data necessitate transmission, storage, and offline processing using conventional computers or artificial neural networks (ANNs). This process is power-intensive, slow, and prevents real-time analysis and feedback. Real-time processing at the recording site is highly desirable for improving neural probe capabilities and understanding nervous systems. Reservoir computing (RC), derived from recurrent neural networks (RNNs), offers a solution. RC uses a dynamic system (reservoir) to perform nonlinear transformations of input signals, projecting them into a high-dimensional space (reservoir states). These states are then processed by a small, trained linear ANN (readout layer). A key feature is fading memory, where the reservoir state depends on recent inputs, making it suitable for analyzing temporal features in neural spike trains. Dynamic memristors, with inherent short-term memory, are promising candidates for RC hardware due to their simple structure and tunable characteristics through material engineering. This research experimentally demonstrates a memristor-based RC system for real-time neural data analysis, using a perovskite halide-based memristor with low switching voltage and current.

The paper cites numerous studies on neural probe technologies and their limitations in handling large datasets for offline analysis. It reviews the application of reservoir computing in various tasks and the use of memristors as reservoirs for temporal data processing. It also discusses the challenges of using existing memristors with high programming voltage and current for direct neural signal processing, and the potential benefits of using perovskite halides with their low activation energies for ion migration.

The researchers developed a perovskite halide-based memristor (Ag/CsPbI3/Ag) with ultralow switching voltage (<100 mV) and current (1-100 nA). The memristor's operation was characterized through current-voltage (I-V) curves, showing volatile memristive effects with an average SET voltage of ~80 mV. Energy-dispersive spectroscopy (EDS) confirmed iodine concentration reduction after SET processes, suggesting iodine vacancy generation as the memristive mechanism. Cyclic voltammetry further supported the electrochemical reactions between silver electrodes and iodine ions. The memristor's response to low-voltage pulses (100 mV, 5 ms, or lower current with reduced pulse width) was analyzed, demonstrating nonlinear current increase and a short-term memory effect with a relaxation time of ~100 ms. The response to emulated action potentials was also studied, showing qualitatively similar responses to simple digital pulses. The memristor's suitability as a reservoir was confirmed through tests demonstrating internal dynamics, nonlinearity, fading memory, separability, and echo state properties. For neural firing pattern recognition, four common patterns (Tonic, Bursting, Irregular, Adapting) were emulated using square-wave pulses (100 mV, 2 ms). A single memristor, with the concept of virtual nodes to expand the reservoir size, was used. Conductance states were sampled every 20 ms, creating virtual nodes reflecting temporal features. These states were fed to a fully connected (FC) neural network (readout layer) for classification. For real-time detection of firing pattern changes, a bilayer convolutional neural network (CNN) was used as the readout layer. For real-time analysis of neural synchronization states, two memristors were used as the reservoir, with the reservoir states fed into a CNN readout layer. Both simulation and experimental results were used to evaluate the performance of the RC system for pattern recognition and synchronization state analysis.

The researchers successfully fabricated a low-voltage, dynamic memristor based on CsPbI3, suitable for direct interfacing with neural signals. The memristor exhibited a short-term memory effect, crucial for processing temporal information in neural spike trains. The memristor-based RC system accurately recognized four common neural firing patterns (Tonic, Bursting, Irregular, Adapting) with an accuracy of ~87%. It also successfully monitored transitions between these patterns in real-time, accurately identifying pattern changes in a streaming spike train. Furthermore, the system successfully analyzed the synchronization states (in-phase, anti-phase, no phase) between two neurons and detected transitions between these states, demonstrating its capability to resolve correlations between spikes from different neurons. Simulations showed the potential to expand this system to analyze synchronization in larger neural networks.

The study successfully demonstrates the feasibility of using a memristor-based RC system for real-time neural activity analysis. The low-voltage memristor design enables direct interfacing with neural signals without preprocessing, a significant improvement over conventional methods. The high accuracy in firing pattern recognition and synchronization analysis highlights the system's potential for neuroscience research and applications. The ability to detect firing pattern transitions in real-time is particularly valuable for studying neural responses to stimuli. The scalability of the system, demonstrated by simulations, suggests its applicability to larger neural networks. However, emulated spike trains were used, and real-world neural signals may exhibit greater complexity that needs to be further investigated.

This work demonstrates a novel memristor-based reservoir computing system for real-time analysis of neural activity. The use of a low-voltage CsPbI3 memristor enables direct processing of neural spike trains, eliminating the need for offline data processing. The system shows high accuracy in recognizing neural firing patterns, tracking their transitions, and detecting neural synchronization. Future research should focus on improving the readout layer, developing advanced RC algorithms, and optimizing memristor hardware for large-scale integration with advanced neural probes to enable real-time analysis of complex interactions within large biological neural networks. The system could also be explored for active interaction with biological networks.

The study primarily used emulated neural spike trains for testing. Real-world neural spike trains are often much more complex and noisy, and the system's performance with such signals needs further investigation. The current system's ability to handle diverse and complex temporal features at different timescales could be limited, necessitating the development of memristors with different relaxation rates. The accuracy of synchronization state detection might be affected by the complexity of neural interactions in larger networks.