Continual Learning Neural Network Weight Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Continual learning methods face the challenge of catastrophic forgetting, where model performance on previous tasks significantly decreases when learning new tasks, due to the phenomenon of catastrophic interference, and existing solutions require iterative pruning and retraining, which is impractical.
Innovation Solution
A method that decouples learning parameters into weights and weight scores, using a binary mask to select optimal subnetworks without explicit pruning, freezing previous task weights, and jointly learning weights and scores to prevent catastrophic forgetting and enable forward transfer.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the model learns new tasks sequentially to improve adaptability, then the ability to learn new concepts is enhanced, but catastrophic forgetting causes significant decrease in performance on previous tasks
Solution Approach 1:
The patent segments the neural network weights into multiple groups or subsets, where each group is associated with specific tasks. By dividing the weight parameters into task-specific segments, the model can selectively update only the relevant weight groups for new tasks while preserving the weight groups for previous tasks, thereby preventing catastrophic forgetting while maintaining adaptability to new concepts
Solution Approach 2:
The patent applies local quality by allowing different parts of the neural network (specific weight parameters) to have different update rules. Some weights are designated as task-specific and are updated only for their associated tasks, while other weights serve multiple tasks and are updated more conservatively. This localized differentiation enables the model to adapt to new tasks without uniformly overwriting knowledge from previous tasks
2Reliability
If iterative pruning and retraining is performed to prevent catastrophic forgetting, then knowledge retention improves, but computational complexity and training time increase significantly
Solution Approach 1:
The patent performs preliminary action by pre-designating and organizing weight parameters into task-specific groups before the continual learning process begins. This pre-structuring of weights into manageable subsets allows for efficient selective updates during learning, eliminating the need for computationally expensive iterative pruning and retraining while maintaining effective knowledge retention across tasks
3Productivity
If all weights are updated during learning to improve learning speed, then training efficiency increases, but catastrophic forgetting occurs due to interference between tasks
Solution Approach 1:
The patent extracts and isolates task-specific weight parameters from the overall set of neural network weights. By separating out the weights that are critical for specific tasks and designating them as frozen or protected, the model can rapidly update the remaining non-critical weights for new task learning without causing interference to the extracted task-specific weights, thus achieving both fast learning and stable task performance
Data Source
AI summary
A computing device performs a continual learning method of learning a plurality of task in a sequential order. The computing device uses, in a forward pass of a neural network for learning a current task of the plurality of tasks, a plurality of weights including selected weights, the selected weights being selected in a previous task of the plurality of tasks, freezes the selected weights and updates weights excluding the selected weights from the plurality of weights in a backward pass of the neural network for learning the current task, obtains a binary mask for selecting some weights of the plurality of weights based on a weight score of each of the plurality of weights, and finds a subnetwork of the neural network for the current task based on the binary mask.