Model Updating by Forgetting Low-Frequency Training Regions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing models fail to improve inference accuracy when there is a change in the overall state of the inference target due to insufficient updating with new learning data.
Innovation Solution
A model update device that includes an acquisition unit to specify areas in the explanatory variable space where learning data frequency is low, and updates the model by forgetting or relearning data from these areas to improve inference accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If additional learning is performed using new learning data, then the model can adapt to new patterns, but the inference accuracy may not improve when there is a change in the overall state of the inference target
Solution Approach 1:
The patent applies the principle of discarding and recovering by selectively forgetting outdated learning data that no longer reflects the current state of the inference target. The specification unit identifies and removes learning data from areas with low acquisition frequency, while the update unit performs additional learning with new data. This discarding of obsolete information and recovery through selective relearning enables the model to adapt to state changes and improve inference accuracy.
2Loss of information
If the model is updated by additional learning with all new learning data, then the model incorporates new information, but it retains outdated data that reduces inference accuracy
Solution Approach 1:
The patent applies the principle of taking out (extraction) by extracting and removing specific outdated learning data from the model's knowledge base. The specification unit extracts learning data belonging to areas with low acquisition frequency, and the update unit removes this extracted data before performing additional learning. This selective extraction ensures that only relevant, high-frequency data is retained, improving inference accuracy by eliminating outdated information.
3Measurement precision
If the model forgets learning data from low-frequency areas, then the model reduces outdated information, but it may lose potentially useful data
Solution Approach 1:
The patent applies the principle of parameter changes by using the acquisition frequency of learning data as a dynamic parameter to determine what data to forget. The specification unit changes the parameter of data selection from uniform retention to frequency-based selective retention. By setting a frequency threshold and forgetting data below this threshold, the system dynamically adjusts which data to retain based on its usefulness, improving inference accuracy while minimizing loss of potentially useful information.
Data Source
AI summary
An acquisition unit (11) acquires an explanatory variable that is to be input to a model (37) configured to output an objective variable for the explanatory variable, a specification unit (12) associates a frequency at which an explanatory variable included in each of a plurality of areas, which are obtained by dividing an explanatory variable space, is acquired by the acquisition unit (11) with each of the plurality of areas, and specifies an area to which an explanatory variable included in learning data used to learn the model (37) belongs and in which a frequency of an explanatory variable acquired by the acquisition unit (11) is a predetermined value or less, and an update unit (14) updates the model (37) in such a manner that learning data including an explanatory variable belonging to an area specified by the specification unit (12) is forgotten.