Brain Waves on epilepsy-related research by building and testing multimodal large models,” said Professor Wong. Other new steps include extending the work to invasive neural recordings such as spike trains and electrocorticography, which could provide higher-quality signals for even better adaptation. They are also scaling up the memristor chip architectures and optimising the learning algorithms for faster convergence. Additionally, they are developing highly integrated memristor chips that combine neural signal acquisition, decoding and feedback in a single system. This integrated approach aims to comprehensively improve brain-computer interface performance and establish a stronger foundation for present and future applications. “We encoded 12 different flight commands including take-off, landing, hovering and directional movements along these four degrees of freedom,” said Dr Liu. “This represents a sophisticated control task that demonstrates the practical capability of our memristor-based BCI for complex real-world applications. The successful completion of a 3D flight trajectory around obstacles using only brain signals shows the precision and reliability our system can achieve in demanding control scenarios.” Throughout the sessions, the researchers monitored for errorrelated potentials to trigger coevolutional updates. They tracked both how the memristor decoder parameters evolved and how participants adapted their neural control strategies through multiple update cycles. The team are about to begin a collaboration with Queen Mary Hospital (QMH) to work on epilepsy data and are in the process of getting ethics approval from the Institutional Review Board, a joint body between HKU and the Hospital Authority. “Once this is cleared, we will discuss and collect in-house electroencephalogram datasets from QMH for potential clinical applications of our adaptive neural decoding technology. The collaboration would focus The work is significant in terms of assistive technologies and neurological rehabilitation because the co-evolutionary capability of the system directly addresses one of the biggest barriers to practical BCI deployment: signal drift and variability over time. In rehabilitation settings, where patients’ neural patterns change continuously during recovery, current systems require frequent manual recalibration by technicians. “Think of ‘co-evolution’ as like learning to dance with a partner,” explained Professor Wong, “Both dancers gradually learn each other’s style and adapt their movements to work better together. In our system, the human brain and the memristor decoder are like dance partners learning to collaborate.” “In technical terms, when the system makes a wrong decision, your brain automatically generates a detectable error signal called an error-related potential. Our memristor array monitors for these brain error signals. When detected, we apply small electrical pulses to the memristor hardware to change its resistance, which updates how the decoder interprets your brain patterns,” explained Dr Liu. “Meanwhile, you learn to adjust your mental control strategies based on the system’s feedback. Over time, both your brain and the hardware learn to work together more effectively. Our experiments show this creates an adaptive interaction where both sides contribute to improved orchestration,” explained Professor Wong. The system maintains performance autonomously through hardwarelevel adaptation, with experimental validation showing sustained accuracy over six-hour sessions and approximately 20 per cent improvement compared to static decoders. The ultra-low energy consumption enables extended daily usage without frequent battery replacement. “This could be transformative for patients requiring long-term neural monitoring or those using BCIs for daily assistance, as the system evolves with their changing neural patterns rather than becoming less effective over time,” said Professor Wong. To test the system, 10 healthy participants used brain signals to control drone movements in real time, each completing approximately six hours of testing across multiple sessions. These included controlling the drones at ‘four degrees of freedom’, which refers to the drone’s ability to move in four independent directions – forward and backward, left and right, up and down, and rotational (clockwise and counterclockwise spinning) – in a 3D space. Co-evolutionary capability Control scenarios The key innovation in this research is to showcase that memristor devices can accomplish real-time co-evolution between brain signals and hardware decoders. In simple terms, it is like creating a learning partnership where both the user’s brain and the memristor-based system adapt together, rather than forcing one to accommodate the other. This groundbreaking work represents a multi-institutional strategic collaboration between research teams at HKU, Tsinghua University and Tianjin University, with Professor Wong Ngai and Dr Zhengwu Liu serving as the lead contributors from HKU’s Department of Electrical and Electronic Engineering. “We leverage the memristor’s intrinsic plasticity to implement a co-evolutional process,” said Dr Liu. “When the brain generates error-related potentials following incorrect classifications, these signals trigger direct conductance changes in the memristor array, effectively updating the decoder parameters. Simultaneously, users gradually refine their neural control strategies based on system feedback.” Their approach further consolidates traditional multi-step decoding into single matrix operations, reducing computational complexity significantly while achieving much lower energy consumption compared to conventional CPU-based systems. “For memristor-based brain-computer interfaces (BCIs), this establishes a new paradigm where decoding hardware components serve as active learning partners rather than passive memory and computing elements, addressing the fundamental challenge of signal variability in neural interfaces,” said Professor Wong. Engineering researchers have implemented memristor-based neuromorphic decoders for brain-computer interfaces, creating a groundbreaking decoding system that can effectively co-evolve with changing brain signals. HKU Bulletin | Nov 2025 Research 16 17 This could be transformative for patients requiring long-term neural monitoring or those using braincomputer interfaces for daily assistance, as the system evolves with their changing neural patterns rather than becoming less effective over time. Professor Wong Ngai
RkJQdWJsaXNoZXIy ODI4MTQ=