Data Generation Model for ML Accuracy Deterioration Analysis

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current techniques fail to specify the cause of accuracy deterioration in machine learning models due to domain shifts, making it difficult to address the issue effectively.

Innovation Solution

A data generation program that uses a trained data generation model to interpolate feature amounts and losses between datasets from different domains, generating data that approximates the identification loss, allowing for the specification of accuracy deterioration causes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of information

If conventional techniques are used to present wrong data and visualize domain shift, then data on which the model has made mistakes can be presented, but the cause of accuracy deterioration cannot be specified

Engineering Contradiction:
Improvecause specification capabilityVSAvoidaccuracy deterioration analysis precision
Core Design Contradiction:
Loss of informationVSMeasurement precision

Solution Approach 1:

A data generation model is introduced as an intermediary to generate synthetic data that bridges the gap between training data and test data distributions. This intermediary model, trained with both domain-specific loss and identification loss, produces intermediate domain data that reveals the transition path and causes of accuracy deterioration, enabling precise specification of the problems.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent replaces conventional visualization methods with a machine learning-based data generation system. Instead of simply visualizing domain shifts, the system uses a trained generation model to synthetically produce data that systematically reveals the causes of accuracy deterioration through controlled generation processes.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Reliability

If a model is trained only on training domain data, then the model achieves good performance on training data, but the model makes mistakes on test data from different domains

Engineering Contradiction:
Improvemodel performance stabilityVSAvoiddomain adaptation capability
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent changes the training parameters of the data generation model by incorporating identification loss in addition to domain-specific loss. This parameter change enables the model to learn features that are invariant across domains while maintaining domain-specific characteristics, thereby improving both reliability and adaptability.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The data generation model is designed to serve multiple functions: it generates domain-specific data for maintaining specialization while simultaneously learning universal features through identification loss that generalize across domains. This multi-functionality resolves the contradiction between reliability and adaptability.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3996001A1Data generation program, data generation method, and information processing device
Publication Date: 2022.05.11 FUJITSU LTD
  • EP3996001A1 patent drawingFigure 1
  • EP3996001A1 patent drawingFigure 2A~2B
  • EP3996001A1 patent drawingFigure 3

AI summary

A data generation program that causes at least one computer to execute a process, the process includes, acquiring a data generation model that is trained by using a first dataset corresponding to a first domain and a second dataset corresponding to a second domain, and that includes an identification loss by an identification model in a parameter; inputting first data corresponding to the first domain to the identification model to acquire a first identification loss, and inputting second data corresponding to the second domain to the identification model to acquire a second identification loss; generating data in which the second identification loss approximates the first identification loss, by using the data generation model; and outputting the data that is generated.