Continuous learning method, system and equipment based on dynamic expansion prompt generator and storage medium
By adopting a dynamic extension prompt generator in the continuous learning method, retaining previous knowledge and using pseudo task identifiers to activate task-specific information, the problem of forgetting old knowledge when acquiring new knowledge in the existing technology is solved, and more efficient continuous learning performance is achieved.
CN119647540AActive Publication Date: 2025-03-18NANJING UNIV
7 Cites 1 Cited by
Patent Information
- Application Number
- CN202411525281.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-29
- Publication Date
- 2025-03-18
- Estimated Expiration
- 2044-10-29
AI Technical Summary
Technical Problem
Existing continuous learning methods require updating parameters shared across all tasks when acquiring new knowledge, resulting in the forgetting of previously learned knowledge.
Method used
Using a continuous learning method based on the dynamic extension prompt generator, dynamic extension prompt generators are used to reduce catastrophic forgetting by retaining previous knowledge and activating task-specific information using pseudo task identifiers.
Benefits of technology
Effectively reduce catastrophic forgetting, promote seamless acquisition of new tasks, and greatly promote the improvement of continuous learning performance.
✦ Generated by Eureka AI based on patent content.
Abstract
The invention discloses a continuous learning method, system and device based on a dynamic expansion prompt generator and a storage medium. The method comprises the steps that a data set is divided into data of different tasks according to continuous learning setting, and continuous learning data is obtained; initializing a prompt-based continuous learning network, and training a model by using data of a first task; performing parameter updating according to a model obtained by training the first task data, and expanding a module of the generator; keeping old lifting generator parameter freezing, and training the updated model by using new task data; repeating the model updating and parameter training process, mapping the image features into pseudo task identity serial numbers, and assisting the model to select corresponding parameters; the system comprises a training module, a parameter updating module, an optimization module and a mapping module. According to the method, disastrous forgetting is effectively reduced, seamless acquisition of new tasks is promoted, and improvement of continuous learning performance is promoted to a great extent.
Need to check novelty before this filing date? Find Prior Art