Learning Apparatus Adjusting Training Data Distribution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models face low inference accuracy due to the predominance of initial training data, which is not adequately improved by re-learning with new supervised data, leading to biased performance.
Innovation Solution
A learning apparatus that adjusts the training set by modifying the distribution of feature values in the training data to include a balanced representation of difficult and easy-to-restore image quality examples, allowing for more comprehensive learning and improved inference accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If re-learning is performed by adding new supervised data to the initial training set, then the learned model can be updated with new information, but the performance is still greatly influenced by the predominant initial training data, resulting in insufficient improvement of inference accuracy
Solution Approach 1:
The training set is segmented into two distinct parts: initial training data and newly-added supervised data. By separating these data sources and applying different weighting or selection strategies to each, the system can control the influence of initial training data while incorporating new information, thereby improving inference accuracy without being overly dominated by the original training distribution.
Solution Approach 2:
The patent changes the parameter of data distribution by adjusting the proportion or weight of newly-added supervised data relative to initial training data. This parameter adjustment allows the model to adapt to new information more effectively while reducing the predominant influence of initial training data, thus resolving the contradiction between maintaining reliability and improving adaptability.
2Productivity
If a large amount of initial training data is used to train the learned model, then the model can achieve basic functionality, but the inference accuracy remains low when the training data distribution is biased or incorrect
Solution Approach 1:
The patent extracts and identifies the problematic portions of the initial training data that cause biased performance. By separating out the useful initial training data from the harmful or biased portions, and selectively combining them with newly-added supervised data, the system can maintain training efficiency while improving inference accuracy by removing the negative influence of biased data.
Solution Approach 2:
Instead of simply adding new data to the existing training set, the patent inverts the approach by selectively replacing or reweighting portions of the initial training data that cause biased performance. This inversion allows the model to learn from new supervised data more effectively while discarding or reducing the influence of harmful initial training data, thereby improving inference accuracy without sacrificing training efficiency.
Data Source
AI summary
A learning apparatus includes an adjustment unit configured to, for a training set including a plurality of pieces of training data, adjust the number of pieces of training data included in the training set such that feature values of the plurality of pieces of training data have a predetermined distribution, and a training unit configured to perform machine learning using the training set to generate a learned model.