Continual Learning Model Parameter Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for continual learning of neural networks face challenges such as forgetting previously learned knowledge, computational inefficiency, and the need for access to large training datasets when adapting to new tasks.
Innovation Solution
A computer-implemented method for continual learning that involves providing a pre-trained model for a first task, selecting specialized parameters for a second task, and adapting the model by updating only these specialized parameters, while freezing other parameters to maintain knowledge from previous tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If all parameters of the model are updated during continual learning, then the model can adapt to new tasks, but previously learned knowledge is forgotten
Solution Approach 1:
The patent segments the model parameters into two distinct groups: core parameters that remain frozen to preserve previously learned knowledge, and specialized parameters that are updated to adapt to new tasks. This segmentation allows selective updating of only the necessary parameters, preventing catastrophic forgetting while maintaining adaptability to new tasks.
2Adaptability or versatility
If extra parameters are added to the model to facilitate learning of new information, then the model can adapt to new tasks, but determining the optimal set of parameters becomes challenging and expanding the model becomes cumbersome
Solution Approach 1:
The patent extracts and isolates only the specific parameters that are necessary for adapting to new tasks, rather than adding comprehensive extra parameters. By identifying and updating only the specialized parameters that have the greatest impact on new task performance, the method avoids the complexity of determining optimal parameter sets and prevents cumbersome model expansion.
3Adaptability or versatility
If the model is adapted to new tasks by updating parameters, then the model can learn new information, but computational overhead increases
Solution Approach 1:
The patent applies partial action by updating only a subset of parameters (the specialized parameters) rather than all parameters in the model. This selective updating significantly reduces computational overhead and energy consumption while still achieving effective adaptation to new tasks, as only the parameters most relevant to new task performance are modified.
4Loss of information
If the model is pre-trained on vast datasets, then the model gains broad knowledge, but the model lacks specific knowledge for specific tasks
Solution Approach 1:
The patent uses preliminary action by first pre-training the model on vast diverse datasets to establish broad knowledge and powerful generic capabilities. Then, when specific task knowledge is needed, the model is fine-tuned by updating specialized parameters on task-specific data. This two-stage approach ensures the model retains its broad knowledge while acquiring specific task expertise.
Data Source
Figure 1
Figure 2
Figure 3A~3B
AI summary
A computer-implemented continual learning method is provided. The method comprises steps of: (S10) providing a model pre-trained for a first task; (S20) selecting specialized parameters of the model that are specialized for a second task; and (S30) adapting the model to the second task by updating said specialized parameters.