Machine Learning Model Training with Human Vision Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning models for image data processing do not effectively incorporate human vision mechanisms, particularly when dealing with components difficult to visually judge or noise components that are hard to distinguish visually, leading to suboptimal recognition and generation performance.
Innovation Solution
The proposed learning method involves processing image data by either omitting components difficult to visually judge or adding noise components that are difficult to visually judge at a predetermined ratio, and using these modified data sets to train machine learning models for classification and generation tasks, including Generative Adversarial Networks (GANs).
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data augmentation is performed to increase training data quantity, then recognition performance is improved, but the model learns features that are not visually significant
Solution Approach 1:
The patent applies local quality by differentiating between visually significant components and visually insignificant components (noise) in image data. The learning model is designed to selectively attend to and learn from visually significant features while ignoring or downweighting visually insignificant noise components. This is achieved through mechanisms such as attention maps, feature importance weighting, or hierarchical processing that allows different regions and features to have different weights in the learning process.
2Reliability
If noise components are added to training data to improve robustness, then adversarial robustness is improved, but visual quality of generated images deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-processing training data to separate and identify noise components before the main learning process. The system performs preliminary steps such as denoising, thresholding, or feature extraction to remove or isolate visually insignificant components from the training data. This preliminary action ensures that the learning model receives cleaned data that maintains both robustness and visual quality, preventing noise from degrading generated image quality while still providing robustness benefits.
3Measurement precision
If human vision mechanism is incorporated in the model, then recognition accuracy is improved, but model complexity increases
Solution Approach 1:
The patent applies the intermediary principle by introducing an intermediate layer or module that bridges the gap between raw image data and the final recognition task. This intermediary component (such as an attention mechanism, feature selection layer, or visual hierarchy module) processes and filters information to align with human vision mechanisms. It acts as a mediator that simplifies the data representation into visually meaningful features, improving recognition accuracy while keeping the overall model structure manageable and interpretable.
Data Source
AI summary
A learning method, a learning model, a classifier, a generator, and a processing system are provided, which consider human vision in learning using a machine learning model for an image. The learning method learns a machine learning model that inputs or outputs image data with data for learning that includes both or either one of image data in which a component that is difficult to judge visually is left out and image data in which a noise component that is difficult to judge visually is added at a predetermined ratio.