Neural Network Weight Importance Control for Catastrophic Forgetting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Catastrophic forgetting occurs when neural networks are trained on new tasks, causing a decrease in accuracy for previously learned tasks due to the updating of important weight coefficients.
Innovation Solution
A method that adjusts weight coefficients based on importance parameters and gradients during learning, suppressing updates for important coefficients while allowing some updates to maintain performance on new tasks, using importance parameters to determine the update amount.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If weight coefficients are updated during learning of new tasks, then the neural network can learn new tasks, but accuracy for previously learned tasks decreases due to catastrophic forgetting
Solution Approach 1:
The patent applies local quality by differentiating the treatment of weight coefficients based on their importance to previously learned tasks. Important weight coefficients (those with high importance parameters) are protected from updates, while less important ones are allowed to update. This selective update approach based on local importance characteristics resolves the contradiction by preserving critical knowledge while allowing adaptation in less critical areas.
Solution Approach 2:
The patent implements dynamics by making the update rule for weight coefficients adaptive rather than static. The importance parameter dynamically determines the degree of protection for each weight coefficient, and this importance is recalculated based on gradients from previously learned tasks. This dynamic adjustment allows the system to balance between preserving past knowledge and learning new tasks.
2Reliability
If weight coefficients important for previously learned tasks are stopped from being updated, then catastrophic forgetting is suppressed, but transfer learning is inhibited
Solution Approach 1:
The patent applies partial action by allowing some updates to important weight coefficients rather than completely preventing updates. The update amount is controlled by the importance parameter, which scales the gradient update. This partial update approach prevents catastrophic forgetting while still allowing some adaptation for transfer learning, resolving the contradiction between preserving past knowledge and enabling new learning.
Solution Approach 2:
The patent changes the parameter (importance parameter) that controls the update behavior of weight coefficients. By introducing this new parameter that reflects the importance of each weight coefficient to previously learned tasks, the system can dynamically adjust the degree of protection, enabling both catastrophic forgetting suppression and transfer learning facilitation through parameter-based control.
3Productivity
If all weight coefficients are updated uniformly during new task learning, then learning efficiency is high, but important previously learned knowledge is lost
Solution Approach 1:
The patent applies local quality by treating different weight coefficients differently based on their individual importance parameters. Instead of uniform treatment, each weight coefficient receives a customized update magnitude proportional to its importance. This selective approach maintains learning efficiency for less important parameters while protecting important knowledge, resolving the contradiction between efficiency and knowledge preservation.
Solution Approach 2:
The patent implements feedback by calculating importance parameters based on gradients from previously learned tasks and using this information to modulate updates during new task learning. The importance parameter acts as a feedback signal that informs the update process about which weight coefficients are critical, allowing the system to adjust its learning behavior accordingly to prevent knowledge loss.
Data Source
AI summary
The neural network includes layers, and the layers each include a plurality of units, the plurality of units each have a weight coefficient associated with each input to a unit, and an importance parameter indicating importance of the weight coefficient. The method includes: in learning of a task, adjusting a first weight coefficient based on the importance parameter of the first weight coefficient and a first gradient of the first weight coefficient determined using a training set; and after the learning has been completed, determining, based on the training set, respective second gradients of a plurality of second weight coefficients included in a first layer including the first weight coefficient, and calculating the importance parameter of the first weight coefficient to be used in learning of a next task based on the respective second gradients.