Sample-Specific Data Augmentation Parameter Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data augmentation methods for machine learning models, such as those described in Patent Literature 1, do not effectively design optimal data augmentation parameters for individual object samples, leading to suboptimal generalization of the target value prediction model.
Innovation Solution
A machine learning apparatus that includes an acquisition unit, a data augmentation unit, and a first training unit, which acquires a training sample, generates a first augmented sample through data augmentation, and trains a parameter output function to output a second data augmentation parameter specific to each object sample, thereby enhancing data augmentation for individual samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data augmentation is applied using fixed or common parameters for all samples, then the implementation is simple, but the generalization performance of the target value prediction model is suboptimal
Solution Approach 1:
The system enables each object sample to generate its own optimal data augmentation parameters through the parameter output function, eliminating the need for manual expert design. The model learns to automatically determine appropriate transformation parameters (rotation angle, enlargement ratio, etc.) for each individual sample based on its characteristics, achieving sample-specific optimization without external intervention.
Solution Approach 2:
The invention transforms fixed or common data augmentation parameters into dynamic, sample-specific parameters. The parameter output function generates different augmentation parameters for each object sample, allowing the system to adapt transformation intensity and type to individual sample characteristics, thereby improving generalization performance while maintaining automated operation.
2Reliability
If sample-specific data augmentation parameters are designed manually by experts, then the optimization for individual samples is achieved, but the process requires high-level expertise and is time-consuming
Solution Approach 1:
The system replaces manual expert design with automated machine learning-based parameter generation. The parameter output function, trained on diverse samples, automatically determines optimal augmentation parameters for new samples without requiring expert intervention, significantly reducing the time and expertise needed while maintaining or improving optimization quality.
Solution Approach 2:
The system performs preliminary training of the parameter output function using a dataset of object samples and their optimal parameters. This pre-trained model then quickly generates appropriate parameters for new samples during deployment, eliminating the need for time-consuming manual expert analysis for each new sample while preserving sample-specific optimization.
3Adaptability or versatility
If diverse data augmentation transformations are applied, then the diversity of training data is enhanced, but the risk of losing target value information increases
Solution Approach 1:
The system dynamically adjusts data augmentation parameters based on each sample's characteristics and the desired level of transformation. The parameter output function learns appropriate rotation angles, enlargement ratios, and other transformation parameters that maximize data diversity while staying within bounds that preserve the essential target value information, achieving a balanced optimization.
Solution Approach 2:
The system uses feedback from the training process to refine the parameter output function. By monitoring how different augmentation parameters affect both data diversity and target value preservation, the model learns to generate parameters that achieve the right balance, ensuring diverse transformations do not compromise the essential information needed for accurate prediction.
Data Source
AI summary
According to one embodiment, a machine learning apparatus includes processing circuitry. The processing circuitry acquires a training sample including an object sample and a target value correlated with the object sample. The processing circuitry generates a first augmented sample by applying data augmentation to the object sample in accordance with a first data augmentation parameter. The processing circuitry generates a parameter output function that inputs therein the object sample and outputs a second data augmentation parameter corresponding to the object sample, by machine learning based on the object sample, the target value and the first augmented sample.