Real-Time Training System for Large Models Based on Human-Brain Mechanisms

The invention discloses a real-time training system for large models based on human-brain mechanisms. The system comprises: an information input module for receiving multi-type data and standardizing it; a temporary storage module for short-term buffering; an importance-evaluation module for real-time scoring of buffered inputs; a memory-classification storage module that stores data into short-, mid-, and long-term memory according to importance; a real-time fine-tuning/training module that updates parameters via LoRA with small batches; and a model-update control module that monitors performance and automatically rolls back upon anomalies. By introducing real-time fine-tuning and brain-inspired memory/forgetting, the system improves response speed and accuracy while reducing resource consumption and adapting to dynamic environments. CN202510456395_ABS_FT

This record disseminates, with a citable DOI, the technical content and supporting materials of the Chinese patent CN202510456395A (publication date: 2025-08-01; applicant: Zhejiang University). The invention discloses a real-time training system for large models based on human-brain mechanisms. The system includes multi-type input and preprocessing, working-memory-like buffering, multi-factor importance evaluation, hierarchical memory classification, LoRA-based real-time fine-tuning, and safety rollback control. By introducing brain-inspired memory/forgetting and real-time adaptation, it enhances responsiveness, accuracy, and resource efficiency under dynamic conditions.