Conditional GAN Verification for Designated-Class Output Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing conditional generative adversarial networks (CGANs) face challenges in accurately generating data belonging to a designated class, as the quality of the generator can lead to misclassification by classification models.
Innovation Solution
An information processing device that includes learning a generative model and a classification model, with formal verification to ensure the generative model does not produce outputs classified into a class different from the designated class within a certain norm of the noise vector input.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a conditional GAN is used to generate data of a designated class, then the ability to control the generated data class is improved, but the accuracy of generating data belonging to the designated class deteriorates due to generator quality limitations
Solution Approach 1:
The patent applies preliminary action by performing formal verification before the generative model is deployed for data generation. The verification process checks whether the model satisfies the property that outputs are correctly classified within a certain norm range of noise vectors, preventing misclassification issues before they occur in practical application.
Solution Approach 2:
The patent substitutes the traditional empirical evaluation mechanism with a formal verification mechanism. Instead of relying on trial-and-error or statistical testing to ensure classification accuracy, the system uses mathematical formal verification to rigorously prove that the generative model satisfies the desired classification property within specified noise ranges.
2Manufacturing precision
If formal verification is applied to ensure correct classification, then the accuracy of generating data of the designated class is improved, but the complexity of the system increases
Solution Approach 1:
The patent extracts the verification function as a separate, independent module from the generative model itself. The formal verification process operates as an external tool that analyzes the trained model without modifying its core structure, thereby adding verification capability while minimizing integration complexity.
Solution Approach 2:
The patent manages complexity by allowing users to adjust the norm parameter that defines the range of noise vectors to be verified. This parameter can be tuned to balance between verification thoroughness and computational complexity, enabling flexible control over the verification scope based on specific application requirements.
Data Source
AI summary
An information processing device executes learning a generative model that generates data belonging to a designated class on the basis of a noise vector and the designated class by executing a learning algorithm of a conditional generative adversarial network, learning a classification model that classifies input data based on whether the input data is in the designated class, and determining whether a property is satisfied when the classification model classifies an output of the generative model by executing a formal verification algorithm. The property indicates that the generative model does not generate data classified into a class different from a first class designated for the generative model, within a range of a certain norm of a noise vector input to the generative model.