一种基于双层联合约束优化的GIS大语言模型持续学习方法
By employing a two-layer joint constraint optimization strategy, the problems of knowledge fusion and forgetting in the continuous learning of the GIS large language model were solved, realizing multi-stage continuous fusion and stable representation of GIS professional knowledge, and improving the model's adaptability and stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA UNIV OF MINING & TECH
- Filing Date
- 2026-05-15
- Publication Date
- 2026-07-17
AI Technical Summary
Existing technologies struggle to enable continuous learning of large language models in the GIS field, particularly in maintaining existing knowledge while avoiding catastrophic forgetting, and in effectively integrating multi-stage GIS knowledge.
A two-layer joint constraint optimization strategy is adopted. The pre-trained model is aligned and trained using the low-rank adaptive LoRA algorithm. Instantaneous gradient direction constraints and global parameter drift constraints are constructed to enable the model to continuously integrate and stably retain multi-stage GIS professional knowledge in a unified parameter space.
It enhances the model's continuous learning ability, adapts to the dynamic evolution of application needs in the GIS field, ensures rapid absorption of new knowledge and complete retention of historical knowledge, and improves the model's stability and adaptability.
Smart Images
Figure CN122198038B_ABST