Continual Learning Buffer Updates to Prevent Catastrophic Forgetting

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing AI-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 defect diagnosis using continual learning that selects and updates training data by calculating the degree of influence and change in models, utilizing a buffer training data set to maintain previous knowledge while adapting to new data without excessive storage.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If all previous training data is relearned to prevent catastrophic forgetting, then accuracy of previous data is maintained, but storage space requirements increase excessively

Engineering Contradiction:
Improveaccuracy of previous dataVSAvoidstorage space
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent extracts only the most critical training data points from the historical dataset and stores them in a buffer memory. By calculating importance scores for each data point and selectively retaining only those above a threshold, the system prevents catastrophic forgetting while using minimal storage space, rather than storing all historical data

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the parameter of data retention from binary (keep all or keep none) to a continuous importance score-based selection. By introducing an importance calculation mechanism that assigns scores to different data points, the system can dynamically adjust which data to retain in the buffer, optimizing both accuracy maintenance and storage efficiency

Inventive Principle:
Principle #35Parameter changes

2Quantity of substance

If a buffer memory is introduced to store selected training data, then storage space is reduced, but device complexity increases

Engineering Contradiction:
Improvestorage spaceVSAvoidsystem structure
Core Design Contradiction:
Quantity of substanceVSDevice complexity

Solution Approach 1:

The patent implements self-service by having the system automatically calculate importance scores for training data points and autonomously select which data to store in the buffer. The importance calculation unit and selection mechanism operate without external intervention, reducing the need for complex manual data management structures while achieving efficient storage

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent applies discarding and recovering by systematically evaluating training data points, discarding those with low importance scores from buffer storage, and recovering only the essential data points needed for preventing catastrophic forgetting. This selective approach simplifies the storage system compared to retaining all data

Inventive Principle:
Principle #34Discarding and recovering

3Reliability

If importance calculation and buffer updating mechanisms are implemented, then catastrophic forgetting is minimized, but training time and computational resources increase

Engineering Contradiction:
Improvemodel performance stabilityVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent applies partial action by calculating importance scores and selecting only a subset of training data points for buffer storage, rather than processing or storing all historical data. This partial processing approach significantly reduces computational overhead and training time while still effectively preventing catastrophic forgetting through selective data retention

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent implements preliminary action by pre-calculating and storing important training data points in the buffer during offline phases. This preliminary preparation allows the system to quickly access critical data during online inference without performing complex calculations in real-time, reducing operational training time

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4657327A1Device for updating training data sets for continual learning and method therefor
Publication Date: 2025.12.03 DOOSAN ENERBILITY CO LTD
  • EP4657327A1 patent drawingFigure 1~2
  • EP4657327A1 patent drawingFigure 3
  • EP4657327A1 patent drawingFigure 4

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.