news

China News Service, Beijing, August 17 (Reporter Sun Zifa) In response to the problems faced by the general artificial intelligence (AI) path "based on exogenous complexity", such as unsustainable computing resources and energy consumption and insufficient explainability, the research team of Li Guoqi and Xu Bo from the Institute of Automation of the Chinese Academy of Sciences, in collaboration with peer scholars from Tsinghua University, Peking University and other institutions, drew on the complex dynamic characteristics of brain neurons and proposed a method for constructing brain-like neuron models "based on endogenous complexity" in the latest research.

This new brain-like computing method can improve the computing resource consumption problem caused by the outward expansion of traditional models, and also provides an example for the effective use of neuroscience to develop artificial intelligence. The relevant results paper was recently published in the international professional academic journal Nature Computational Science.

The cooperation team said that building a more general artificial intelligence and giving the model a more extensive and general cognitive ability is an important goal in the current development of artificial intelligence. The current popular large model path is based on the "scaling law" to build larger, deeper and wider neural networks, which can be called a general intelligence implementation method "based on exogenous complexity", but this path faces problems such as unsustainable computing resources and energy consumption and insufficient explainability.

In this study, the collaborative team first demonstrated the equivalence of the dynamic characteristics of the LIF model and the HH model of the spiking neural network neurons, and further proved theoretically that the HH neuron can be equivalent to the dynamic characteristics of four time-varying parameter LIF neurons (tv-LIF) with a specific connection structure.

Based on this equivalence, the team designed a micro-architecture to enhance the intrinsic complexity of the computing unit, enabling the HH network model to simulate the dynamic characteristics of the larger-scale LIF network model and achieve similar computing functions on a smaller network architecture. Subsequently, the team further simplified the "HH model" (tv-LIF2HH) constructed by 4 tv-LIF neurons into an s-LIF2HH model, and verified the effectiveness of this simplified model in capturing complex dynamic behaviors through simulation experiments.

The experimental results of this study show that the HH network model and the s-LIF2HH network model have similar performance in representation and robustness, verifying the effectiveness and reliability of the intrinsic complexity model in dealing with complex tasks. At the same time, the study also found that the HH network model is more efficient in computing resource consumption, significantly reducing the use of memory and computing time, thereby improving the overall computing efficiency.

The collaborative team explained their research results through the information bottleneck theory and believed that this study provides new methods and theoretical support for integrating the complex dynamic characteristics of neuroscience into artificial intelligence, and provides a feasible solution for the optimization and performance improvement of artificial intelligence models in practical applications.

It is revealed that the cooperation team has currently started research on larger-scale HH networks and multi-branch multi-chamber neurons with greater intrinsic complexity, which is expected to further improve the computing efficiency and task processing capabilities of large models and achieve rapid implementation in actual application scenarios. (End)

(China News Network)