Developing more general artificial intelligence with broader and more universal cognitive abilities is an important goal in today’s AI research. The currently popular large-model approach relies on the “Scaling Law” to build larger, deeper, and wider neural networks—an approach that can be described as achieving general intelligence through “exogenous complexity.” This path faces challenges such as unsustainable demands for computational resources and energy, as well as limited interpretability. Researchers Li Guoqi and Xu Bo from the Institute of Automation, Chinese Academy of Sciences, in collaboration with teams from Tsinghua University and Peking University, drew inspiration from the complex dynamics of biological neurons to propose a “brain-inspired neuron model based on endogenous complexity.” This approach addresses the computational resource consumption problem caused by simply scaling up traditional models and offers an example of how neuroscience can inform AI development.

This study first demonstrates the dynamical equivalence between the Leaky Integrate and Fire (LIF) model and the Hodgkin-Huxley (HH) neuron model [1]. It further theoretically proves that an HH neuron can be represented by four time-varying parameter LIF (tv-LIF) neurons with specific connectivity. Based on this equivalence, the team designed micro-architectures that enhance the endogenous complexity of computational units, enabling the HH network model to simulate the dynamics of much larger LIF networks while achieving similar computational functionality in a more compact architecture. Furthermore, the team simplified the “HH model” built from four tv-LIF neurons (tv-LIF2HH) into an s-LIF2HH model and validated through simulations that this simplified model effectively captures complex dynamical behaviors.

Experimental results show that the HH network model and the s-LIF2HH network model have comparable performance in representational capacity and robustness, verifying the effectiveness and reliability of endogenous complexity models for handling complex tasks. Additionally, the study found that the HH network model is more efficient in terms of computational resource usage, significantly reducing memory and computation time, thereby improving overall operational efficiency. The research team also interpreted these findings using information bottleneck theory.

By incorporating complex dynamical properties from neuroscience into AI models, this work provides a new method and theoretical foundation for bridging artificial intelligence and neuroscience. It also offers practical solutions for optimizing and improving the performance of AI models in real-world applications. The team is now working on larger-scale HH networks and more complex multi-branch, multi-compartment neurons with greater endogenous complexity, aiming to further enhance large model computational efficiency and task-handling capabilities for rapid deployment in practical scenarios.

This work was published in Nature Computational Science. Corresponding authors are Professor Li Guoqi and Professor Xu Bo of the Institute of Automation, Chinese Academy of Sciences, and Professor Tian Yonghong of Peking University. Co-first authors include He Linxuan (an undergraduate in Tsinghua University’s Qian Xuesen Class and intern at the Institute of Automation), Xu Yunhui (an undergraduate in the Mathematical Sciences Base Class and intern at the Institute of Automation), and PhD students He Weihua and Lin Yihan from Tsinghua University’s Department of Precision Instrument.

Paper link:https://www.nature.com/articles/s43588-024-00674-9

Comment link:https://www.nature.com/articles/s43588-024-00677-6

[1] The HH neuron model, or Hodgkin-Huxley model, was proposed by British physiologists Alan Hodgkin and Andrew Huxley in 1952 based on electrophysiological data from squid giant axons. It describes the generation and propagation of neural spikes, for which they were awarded the 1963 Nobel Prize in Physiology or Medicine. The model consists of a set of nonlinear differential equations that reflect the opening and closing of ion channels in the cell membrane and their relationship to membrane potential changes. It is a landmark in neuroscience, providing the first molecular-level explanation of action potential generation and laying the foundation for subsequent research in neuronal electrophysiology.