Data Augmentation via Generative Models for Protected Attribute Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models face challenges in estimating protected attributes due to data scarcity and over-training on specific attributes, leading to inadequate generalization and ethical concerns, especially when direct use of such attributes is not possible.
Innovation Solution
A data augmentation method that utilizes generation models to generate images and captions from text queries, extracting additional attributes from captions to augment the training data set, thereby increasing the diversity and reducing over-training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a machine learning model is trained using only the attributes included in the original data set, then the training process is simple and fast, but the model may perform over-training to specific attributes and cannot generalize well to new data
Solution Approach 1:
The patent applies data augmentation before training the machine learning model. Generation models are used to synthesize additional training data with diverse attribute combinations in advance, expanding the training dataset beyond the original attributes. This preliminary expansion of training data prevents over-training while maintaining efficient training processes.
2Adaptability or versatility
If the data set is augmented by adding more attributes, then the model's generalization capability improves, but the data processing complexity and computational cost increase
Solution Approach 1:
The patent introduces generation models as intermediary components that automatically synthesize augmented training data. These generation models act as mediators between the original data set and the machine learning model, generating diverse attribute combinations without requiring manual data processing. This automation reduces the complexity of data processing while achieving effective data augmentation.
3Measurement precision
If protected attributes are directly used in the machine learning model, then the model can accurately estimate protected attributes, but legal and ethical risks arise
Solution Approach 1:
The patent extracts and removes protected attributes from the training process. Instead of directly using protected attributes, the system trains the model to estimate them indirectly through other attributes. The generation models create training data that excludes protected attributes, forcing the model to learn alternative estimation pathways that avoid legal and ethical issues while maintaining estimation capability.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An information processing apparatus acquires an image by inputting, into a first generation model, first text including attribute values of a plurality of attributes included in first data among a plurality of pieces of data. The information processing apparatus acquires second text by inputting the image to a second generation model. The information processing apparatus selects an attribute value of other attribute different from the plurality of attributes from the second text. The information processing apparatus augments a plurality of pieces of data by adding the attribute value of the other attribute to the first data. To augment a data set for training a machine learning model for estimating a protected attribute.