Transfer Learning Domain Adaptation With Synthetic Target Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automated testing equipment (ATE) systems face challenges in efficiently adapting machine learning models to data-sparse target domains due to limited data availability, leading to overfitting and reduced performance in manufacturing environments.
Innovation Solution
The use of smear regularization and Sliced Wasserstein Optimal Transport to align data distributions, combined with moment matching from a data-rich source domain, enables the creation of synthetic data in the target domain, enhancing the robustness and stability of transfer learning for ATE systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional machine learning models are trained on limited target domain data, then training speed is maintained, but model reliability and accuracy deteriorate due to overfitting
Solution Approach 1:
The patent creates synthetic copies of target domain data by transporting source domain data through learned optimal transport maps. These synthetic copies augment the limited target domain training data, enabling more robust model training without requiring additional physical data collection, thereby improving model reliability while working within data quantity constraints
Solution Approach 2:
The patent introduces optimal transport maps as an intermediary mechanism that bridges the source domain and target domain. These maps enable indirect data generation by transforming source domain samples into synthetic target domain samples, allowing the model to learn from augmented data without direct access to large quantities of target domain data
2Measurement precision
If more target domain data is collected to improve model accuracy, then model accuracy improves, but data collection time and cost increase
Solution Approach 1:
The patent performs preliminary domain adaptation by learning optimal transport maps between source and target domains before final model training. This preliminary action creates a transformed feature space where data distributions are aligned, enabling accurate model training with limited target domain data and eliminating the need for extensive subsequent data collection
Solution Approach 2:
The patent generates synthetic target domain data copies from source domain data through optimal transport transformation. These synthetic copies serve as additional training samples, improving model accuracy without requiring time-consuming physical data collection from the target domain
3Adaptability or versatility
If domain adaptation techniques are applied to align data distributions, then transfer learning effectiveness improves, but computational complexity increases
Solution Approach 1:
The patent applies dimensionality reduction techniques to transform high-dimensional source and target domain data into lower-dimensional latent spaces before computing optimal transport maps. This dimensionality change reduces the computational complexity of distribution alignment while preserving the essential structural relationships needed for effective transfer learning
Data Source
Figure 1
Figure 2
Figure 3
AI summary
System and techniques for adapting one domain to another to facilitate transfer learning in a machine learning (ML) model are described herein. Given data from a source domain and a target domain, moments can be mapped from the source domain to the target domain. Synthetic data that is constrained by the mapping can be created in the target domain. The synthetic data can be combined with the target domain data to create training data for a machine learning model to enable the model to be trained to accept target domain data as input and produce an output.