Unsupervised Neural Network for Inverse Problem Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing unsupervised learning methods for inverse problems face challenges in effectively handling unmatched data and lack a systematic mathematical theory to link different approaches such as optimal transport and penalized least squares.
Innovation Solution
The proposed method integrates a cycle-consistent generative neural network with optimal transport theory and penalized least squares, allowing for unsupervised learning of unmatched data by adding a cycle consistency term to the existing structure, assuming a certain probability distribution without fixed measurement assumptions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised deep learning approaches are used for inverse problems, then reconstruction performance is improved, but matched label data must be available which is not available in many applications
Solution Approach 1:
The patent introduces an intermediary framework that connects optimal transport theory with penalized least squares through a cycle-consistent generative neural network. This intermediary structure enables unsupervised learning by mediating between the measurement operator and the reconstruction process, allowing the system to learn from unmatched data without requiring matched label pairs while maintaining reconstruction quality.
Solution Approach 2:
The patent changes the learning paradigm from supervised to unsupervised by modifying the training objective functions. It introduces cycle consistency terms and optimal transport-based loss functions that allow the neural network to learn the inverse problem solution without matched labels, fundamentally changing how the system parameters are optimized during training.
2Adaptability or versatility
If classical PLS, OT, and cycleGAN approaches are used, then unsupervised learning is achieved, but there is no mathematical theory to systematically link these different approaches
Solution Approach 1:
The patent merges optimal transport theory with penalized least squares methodology into a unified mathematical framework. By combining these previously separate approaches through a common cycle-consistent generative neural network structure, the patent creates a systematic theory that links different unsupervised learning methods, reducing theoretical complexity while maintaining versatility.
3Adaptability or versatility
If GAN approaches are used for unsupervised learning, then matched label data is not required, but artificial features are generated due to mode collapsing
Solution Approach 1:
The patent implements feedback mechanisms through cycle consistency constraints that ensure the generated images can be transformed back to the original input. This feedback loop prevents mode collapsing by enforcing that the generative process maintains consistent mappings, thereby improving generation quality and reliability while maintaining unsupervised learning capabilities.
Data Source
AI summary
Disclosed are an unsupervised learning method and an apparatus therefor applicable to general inverse problems. An unsupervised learning method applicable to inverse problems includes receiving a training data set and training an unsupervised learning-based neural network generated based on an optimal transport theory and a penalized least square (PLS) approach using the training data set, wherein the receiving of the training data set includes receiving the training data set including unmatched data.