Learning Model Data Selection for Continual Learning Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current AI models for food image recognition, used to analyze eating habits and calorie counting, suffer from low accuracy and require continual learning due to their inability to adapt to new data, leading to increased training time and redundant data usage.
Innovation Solution
A method and apparatus for operating a learning model that selects and evaluates training data using hard-negative sampling, active-learning, and memory-mapping methods to improve performance, accuracy, and speed, allowing for efficient continual learning by integrating only necessary data into the memory.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training data are continuously added to update the AI model for continual learning, then the classification accuracy for new images is improved, but the training time and update cycle of the AI model gradually increase
Solution Approach 1:
The patent segments the continuous training process into discrete evaluation cycles. The learning model evaluates each new training data item individually, calculating a performance index value for each. Only data items that meet the selection criterion (performance index value greater than threshold) are selected for actual training, thereby segmenting the training workload and reducing overall training time while maintaining accuracy improvements.
Solution Approach 2:
The patent applies local quality by focusing training resources only on specific data items that have high potential to improve model performance. Instead of uniformly training on all new data, the system identifies and selectively trains on data items with performance index values exceeding the threshold, thereby optimizing the quality and efficiency of the training process.
2Adaptability or versatility
If training data are continuously added to update the AI model, then the model adapts to new images, but the amount of redundant data used for training gradually increases
Solution Approach 1:
The patent extracts only the essential and useful training data from the pool of new training data. By calculating performance index values and comparing them against a threshold, the system extracts and selects only those data items that will genuinely contribute to model improvement, discarding redundant data that would not provide meaningful training value.
Solution Approach 2:
The patent applies partial action by training on only a subset of new training data rather than all available data. The selective training approach uses only the portion of data that meets the performance criterion, avoiding the excessive use of redundant data while still achieving effective model adaptation.
3Stability of the object's composition
If the AI model is stuck in the past during learning, then the model maintains stability, but its performance deteriorates over time
Solution Approach 1:
The patent introduces dynamics into the learning model by enabling it to adaptively evaluate and select training data based on current performance needs. The model transitions from a static state where it merely processes all incoming data uniformly to a dynamic state where it actively evaluates each data item's potential impact and selectively learns from high-value data, thereby maintaining both stability and improved performance.
Solution Approach 2:
The patent implements feedback mechanisms where the learning model evaluates new training data by calculating performance index values based on expected improvements. This feedback loop allows the model to identify which data items will most benefit current performance deficiencies, enabling targeted learning that maintains stability while progressively improving performance over time.
Data Source
AI summary
Provided are a method and an apparatus for operating a learning model. A method for operating a learning model according to one embodiment of the present disclosure comprises selecting at least one training data between previous training data and new training data and learning a previous learning model anew using the at least one selected training data.