A Review on Bridging Brain-Inspired Mechanisms and Large-Scale Pre-trained Models: Toward Adaptive, Efficient, and Interpretable AI

作者：Jingzhe Wang, Yajing Wang, Zongwei Luo · 发表于：FinTech and Sustainable Innovation · 年份：2025 · DOI：10.47852/bonviewfsi52026630 · 被引用次数：2 · 研究领域：EEG and Brain-Computer Interfaces、Artificial Intelligence in Healthcare and Education、Functional Brain Connectivity Studies

Large-scale pre-trained models, such as GPT, greatly improve numerous areas of artificial intelligence (AI), including natural language understanding, image recognition, and the integration of various data types. However, these models continue to encounter significant challenges, including excessive computational resource requirements, inadequate adaptability to dynamic environments, susceptibility to catastrophic forgetting, and limited internal interpretability. By comparison, the human brain exhibits efficient learning from sparse data, demonstrates robust adaptability across diverse contexts, operates with minimal energy consumption, retains information over extended periods, and can be elucidated through its underlying cognitive processes. This review examines recent research and is the first to categorize brain-inspired methods into three key dimensions. It explores how mechanisms of the human brain, such as hierarchical and modular designs, biologically inspired attention mechanisms, memory enhancement strategies, synaptic plasticity, and predictive coding, could inspire optimizations for large-scale models. This review not only synthesizes the current state of the field but also proposes potential directions for future research. There remains a need for stronger theories, better testing methods, improved hardware–software coordination, and careful consideration of ethical issues. Overcoming these challenges will require closer collaboration between neuroscientists, co...