Continual Learning Buffer Selection for Accuracy Under Storage Limits
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence-based radiographic testing methods face challenges in maintaining accuracy over time due to catastrophic forgetting, which occurs when relearning all previous data sets, leading to storage space issues.
Innovation Solution
A method and device for diagnosing defects using continual learning, involving training a current model with a buffer training data set and a current data set, calculating influence and change degrees, and updating the buffer data set based on influence and similarity of gradient vectors to minimize storage while maintaining accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all previous training data sets are relearned to prevent catastrophic forgetting, then accuracy of previous data is maintained, but storage space requirements increase excessively
Solution Approach 1:
The patent extracts only the most critical training data from the complete historical data set. A selection unit identifies and extracts high-influence data points based on their impact on model prediction accuracy, storing only these extracted data in a buffer with limited capacity. This resolves the contradiction by maintaining accuracy through selective data retention rather than storing all data.
Solution Approach 2:
The patent changes the parameter of data selection from comprehensive to selective based on influence degree. By calculating and comparing influence degrees of different training data points, the system dynamically selects which data to retain in the buffer, transforming the storage approach from static full-data retention to dynamic selective retention based on accuracy impact.
2Quantity of substance
If a buffer with limited storage capacity is used to store training data, then storage space is reduced, but the ability to maintain prediction accuracy deteriorates
Solution Approach 1:
The patent introduces a new parameter - influence degree - to characterize the importance of training data. By calculating influence degrees based on model prediction differences with and without each data point, the system transforms the buffer selection from random or uniform to optimized based on measured influence, maintaining accuracy with limited storage.
Solution Approach 2:
The patent applies local quality by treating different training data points differently based on their individual influence degrees. High-influence data points receive priority for buffer storage while low-influence points are discarded, creating a non-uniform quality distribution in the stored data that optimizes accuracy for the given storage capacity.
3Quantity of substance
If training data is selectively updated based on influence degree, then storage efficiency is improved, but computational complexity increases
Solution Approach 1:
The patent applies partial action by calculating influence degrees only for a subset of training data points rather than all data. The selection unit processes data incrementally, computing influence metrics for candidate data points and selecting top candidates for buffer storage, reducing overall computational burden while maintaining storage efficiency.
Data Source
AI summary
A method for updating training data sets for continual learning includes providing a buffer with a predetermined storage size, training the current model using a buffer training data set previously stored in the buffer and a current training data set, calculating a degree of change from models trained in at least two previous training rounds of the current model to the current model, determining whether the degree of change is greater than or equal to a reference value, determining an update to the buffer training data set when the degree of change is greater than or equal to the reference value, and updating the buffer training data set based on data points in the current training data set.