Information Processing Control for Incremental Learning Data Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for fine-tuning deep neural network (DNN) models face accuracy degradation due to the tendency to forget initial learning information and are ineffective in handling arbitrarily annotated learning data.
Innovation Solution
An information processing apparatus that manages traceability information of learning models and data, selecting data sets for incremental learning based on evaluation accuracy and similarity to maintain learning accuracy by tracing the history of models and data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If fine tuning is repeated using learning data irrelevant to the initial model, then the model can learn new tasks, but the model forgets information obtained by the initial model
Solution Approach 1:
The system performs preliminary evaluation of candidate learning data sets using the initial model before actual incremental learning. By calculating evaluation accuracy of the initial model on candidate data sets and comparing with reference accuracy, the system pre-screenes data that is likely to cause catastrophic forgetting, thus preventing information loss before it occurs during incremental learning
Solution Approach 2:
The system implements a feedback mechanism where the evaluation accuracy of the initial model on candidate learning data is calculated and compared against reference accuracy. This feedback loop allows the system to iteratively select learning data sets that maintain compatibility with the initial model, thereby preserving previously learned information while enabling new task acquisition
2Ease of manufacture
If learning data include arbitrarily annotated data by users, then data collection becomes easier, but learning accuracy readily decreases
Solution Approach 1:
The system employs the initial model to automatically evaluate and select learning data sets from user-provided arbitrarily annotated data. The initial model serves itself by identifying high-quality learning data without requiring manual verification, thus maintaining ease of data collection while ensuring learning accuracy through automated quality assessment based on evaluation accuracy comparison
3Productivity
If learning data are corrected using existing learned models, then correction efficiency improves, but it becomes impossible to effectively correct/select data with arbitrary annotations
Solution Approach 1:
The system performs preliminary evaluation of candidate learning data sets using the initial model before actual incremental learning. By calculating evaluation accuracy of the initial model on candidate data sets and comparing with reference accuracy, the system pre-screenes data that is likely to cause catastrophic forgetting, thus preventing information loss before it occurs during incremental learning
Solution Approach 2:
The initial model acts as an intermediary between user-provided arbitrarily annotated data and the incremental learning process. Instead of directly using user annotations or relying on previously corrected data, the system uses the initial model's evaluation accuracy as an intermediate metric to objectively assess and select learning data sets, bridging the gap between arbitrary annotations and effective learning
Data Source
AI summary
An information processing apparatus for managing a plurality of learning models, comprises: a model management unit that manages, for each of the plurality of learning models, first information for identifying a data set used to learn the learning model, and second information for identifying an initial model used to learn the learning model; a data management unit that manages third information concerning a plurality of data sets each identified by the first information of each of the plurality of learning models; a reception unit that receives an instruction of incremental learning using, as an initial model, a predetermined learning model included in the plurality of learning models; and a determination unit that determines, based on the first information, the second information, and the third information, a data set to be used for the incremental learning from the plurality of data sets.