The research group is dedicated to leveraging various brain atlas structures, brain functional activities, and brain plasticity mechanisms to develop brain-inspired neural network algorithms that meet the criteria of both biological plausibility and computational efficiency. These algorithms aim for attributes such as low energy consumption, robustness, and flexibility. Our brain-inspired model integrates cutting-edge technologies including deep neural networks, physics or biology-informed neural networks, and graph neural networks topology representation. On the one hand, biologically plausible models enable precise dynamic simulation of complex biological systems such as sleep, anesthesia, and hemodynamics. On the other hand, they can be directly translated into brain-inspired chips for use in specialized application scenarios requiring energy efficiency, stability, and flexibility. These scenarios include, but are not limited to, high-throughput online decoding chips for invasive brain-computer interfaces and high frame rate object recognition chips for dynamic vision sensors. Our research group has introduced several novel learning algorithms for biologically-plausible neural networks, including self-organized backpropagation (Science Advances/IEEE TNNLS), neural modulation-inspired continuous learning (published in Science Advances), dopamine-based reward learning (IEEE TNNLS), membrane potential-based balance tuning (AAAI/IJCAI), neuronal metadynamics (AAAI/NeoComputing), population coding (AAAI), hierarchical network topology algorithms covering multiple brain regions (IEEE TNNLS), neuralODE embedded recurrent neural network (NeuroIPS), and a method for signal alignment of brain signals at different scales (IEEE TCYB). These algorithms have been rigorously tested across various standard AI tasks such as image classification, automatic speech recognition, and natural language processing, as well as continuous action control in certain computer games.

Research Unit:

Brain-inspired algorithm: Integrating biologically plausible mechanisms into conventional artificial neural networks to enhance their overall performance in accuracy, energy efficiency, robustness, flexibility, and other aspects that are challenging for AI but common in biology. These mechanisms encompass neural plasticity, neural circuits, and information encoding. Additionally, our focus lies in developing a general topology extraction method capable of converting any biological circuits or topology into its AI counterpart network to achieve this objective.

Brain-like simulation: Leveraging valuable multiscale findings regarding biological topology, circuits, and neuronal types in pivotal brain regions associated with cognition, such as the prefrontal cortex, hippocampus, and thalamus, to integrate into neural networks. These networks are subsequently employed to simulate brain dynamics, aiding in the improved comprehension of attention, working memory, multisensory integration, and decision-making processes.

Brain-inspired chips: Addressing the demands for interpretability, flexibility, and low energy consumption in some practical applications (e.g., brain-computer interface chip), we have engineered a brain-inspired chip capable of deploying algorithms that mimic the brain with both biological plausibility and computational efficiency. This chip can function as a standalone AI universal chip, supporting dynamic vision, dynamic hearing, and other machine learning tasks, or as a special chip for invasive brain-computer interface.

Mathematical Fundamentals: Concentrating on the mathematical analysis of the brain-inspired algorithms introduced earlier, and delving into the correlation between complex dynamic systems and neural network models. We have introduced fundamental knowledge embedding algorithms, encompassing structures, nodes, loss functions, and operators, to facilitate AI learning for accomplishing tasks such as few-shot learning and long-term prediction.

Smart healthcare: Our focus lies in the development of intelligent anesthesia robots capable of autonomously assessing intraoperative anesthesia depth, pain index, multi-step anesthesia depth prediction, and other indicators using EEG data and additional vital signs. These anesthesia robots leverage brain-inspired reinforcement learning algorithms to make optimal medication decisions, aiding doctors in achieving automated and precise anesthesia and brain state monitoring.