Adversarial Data Augmentation for Debiased Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models, such as image classification models, often learn biased decision rules due to unintended biases in real-world datasets, leading to poor performance in various situations and difficulty in generalizing to different environments.
Innovation Solution
An electronic device is configured to train a biased model using original data and generate adversarial data to create a debiased model, which is then retrained using the adversarial data to reduce bias, thereby improving the model's performance by attacking bias information and generating data without bias labels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a machine learning model is trained using original real-world data, then the model can learn from actual data distributions, but the model learns biased decision rules due to unintended biases in the data
Solution Approach 1:
The patent applies preliminary anti-action by generating adversarial data that deliberately counteracts the biased patterns in original data before the model learns from it. The adversarial data is designed to oppose the biased decision rules, preventing the model from learning incorrect patterns in the first place. This is achieved by creating synthetic data that challenges the biased features while maintaining the underlying true relationships.
Solution Approach 2:
The patent introduces an intermediary mechanism - the adversarial data generation process - that mediates between the biased original data and the model training. This intermediary creates a bridging dataset that preserves useful information from original data while removing harmful biases, allowing the model to learn from a balanced representation.
2Adaptability or versatility
If adversarial data is generated to reduce bias in the model, then the model's generalization improves, but the training process becomes more complex
Solution Approach 1:
The patent applies preliminary action by pre-processing the original data to generate adversarial data before the main model training begins. This adversarial data is prepared in advance and then integrated into the training process, allowing the model to learn from bias-reduced data without requiring complex real-time adjustments during training.
Solution Approach 2:
The patent changes key parameters of the training process by introducing adversarial data with modified characteristics. The adversarial data differs from original data in specific parameters (such as feature distributions or label assignments) to counteract biases, while maintaining other parameters that preserve the true underlying patterns for effective learning.
Data Source
AI summary
A method and electronic device with adversarial data augmentation are provided. The electronic device includes a processor configured to execute instructions; and a memory storing the instructions, where the execution of the instructions by the processor configures the processor to, based on a biased feature within original data, train a biased model to generate biased prediction information using biased training data related to the original data; train a debiased model to generate debiased prediction information using debiased training data, less biased with respect to the biased feature than the biased training data, related to the original data and first adversarial data; and retrain the debiased model using second adversarial data generated based on the biased model and the debiased model.