Synthetic Data Generation for Unsupervised Model Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing unsupervised learning techniques rely heavily on large-scale labeled datasets, which are costly and time-consuming to create, and are hindered by data privacy and usage rights concerns.
Innovation Solution
The proposed solution involves using model-generated data for unsupervised training, where a generative model produces sample images and attention maps based on text prompts, and a target model is trained using these synthetic data and attention maps for image processing tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If large-scale datasets are used for training, then model performance is improved, but data acquisition cost and time consumption increase
Solution Approach 1:
The patent uses a pre-trained generative model to synthesize training data that copies the essential statistical properties and patterns of real-world data. This synthetic data serves as a substitute for collecting and annotating large-scale real datasets, dramatically reducing data acquisition time while maintaining model training effectiveness
Solution Approach 2:
The generative model is pre-trained on real data beforehand to learn data distributions and patterns. This preliminary action enables the model to subsequently generate synthetic training data without requiring time-consuming manual data collection and annotation for each training iteration
2Reliability
If large-scale datasets are used for training, then model performance is improved, but data cost increases
Solution Approach 1:
Synthetic data generated by the pre-trained generative model serves as a cost-effective substitute for expensive real-world datasets. The generative model captures data distributions and generates unlimited synthetic samples without incurring additional data collection, storage, or annotation costs
Solution Approach 2:
The system uses the pre-trained generative model to self-generate training data autonomously without requiring external data sources, human annotators, or expensive data procurement processes, making the training process self-sufficient and cost-effective
3Reliability
If real-world data is used for training, then model performance is improved, but data privacy concerns increase
Solution Approach 1:
The patent generates synthetic data that replicates the statistical properties and patterns of real-world data without containing actual sensitive information. This copying approach maintains model training effectiveness while eliminating privacy risks associated with using real personal or sensitive data
Solution Approach 2:
The pre-trained generative model acts as an intermediary between real-world data and the training process. It transforms real data patterns into synthetic representations that preserve learning value while removing privacy-sensitive information, enabling safe model training
4Quantity of substance
If synthetic data is used for training, then data acquisition cost is reduced, but data richness and diversity may decrease
Solution Approach 1:
The generative model uses text prompts with varying parameters (object types, attributes, relationships, scenarios) to generate diverse synthetic data samples. By systematically changing prompt parameters, the system ensures rich and varied training data coverage without requiring manual data collection
Data Source
AI summary
Embodiments of the disclosure relate to unsupervised learning with synthetic data and attention masks. According to example embodiments of the disclosure, a plurality of sample images are generated by providing a plurality of text prompts into a trained generative model, respectively. For a sample image of the plurality of sample images, at least one attention map is obtained from a generative model, the at least one attention map being determined by the generative model for generating the sample image, an attention map indicating visual elements of an object within the sample image. Training of a target model is performed according to unsupervised learning at least based on the plurality of sample images and attention maps for the plurality of sample images, the target model being configured to perform an image processing task.