Neural Network Domain Translation for Learning Data Generalization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional devices face challenges in generating learning data that effectively improves the generalization performance of neural networks used for estimation tasks, such as class classification and object detection.

Innovation Solution

A learning device comprising an inference unit, a translation unit, and update units that utilize neural networks to translate domain data and update parameters, ensuring the distribution of translated data resembles the original domain data, thereby enhancing generalization performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If conventional devices generate learning data using deep learning techniques (VAE, GAN), then learning data can be synthesized to increase data quantity, but the generalization performance of neural networks for estimation tasks cannot be effectively improved

Engineering Contradiction:
Improvelearning data quantityVSAvoidgeneralization performance
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The patent introduces a translation model as an intermediary component that converts source domain data into target domain data. This translation model acts as a mediator between the source domain and target domain, enabling the generation of learning data that preserves the statistical properties and distribution characteristics of the target domain, thereby improving generalization performance while maintaining adequate data quantity.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent changes the parameters and distribution characteristics of the generated learning data by using domain-specific translation models. Instead of generating data with arbitrary distributions, the system adjusts the data parameters to match the target domain's statistical properties, ensuring that the generated data is suitable for improving generalization performance in estimation tasks.

Inventive Principle:
Principle #35Parameter changes

2Quantity of substance

If neural networks are trained with synthetic learning data, then data quantity can be increased, but the data may not accurately represent the target domain distribution, reducing estimation accuracy

Engineering Contradiction:
Improvelearning data quantityVSAvoidestimation accuracy
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent implements a feedback mechanism where the translation model is trained using paired data from source and target domains, and the generated data is evaluated against the target domain distribution. This feedback loop ensures that the synthetic data continuously improves its representation of the target domain, maintaining both adequate quantity and high estimation accuracy.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent segments the data generation process into distinct components: a translation model for domain conversion, an inference model for estimation tasks, and a training mechanism that uses ground truth labels. This segmentation allows each component to be optimized independently, ensuring that the generated data maintains target domain characteristics while providing sufficient quantity for training.

Inventive Principle:
Principle #1Segmentation

3Quantity of substance

If domain data is translated to generate similar data, then learning data quantity increases, but the complexity of the system increases due to multiple neural networks and update mechanisms

Engineering Contradiction:
Improvelearning data quantityVSAvoidsystem complexity
Core Design Contradiction:
Quantity of substanceVSDevice complexity

Solution Approach 1:

The patent designs the translation model to serve multiple functions: it translates source domain data to target domain data, generates synthetic learning data, and preserves target domain distribution characteristics. This multi-functionality reduces the need for separate specialized components, thereby managing system complexity while achieving increased learning data quantity.

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

Solution Approach 2:

The patent embeds the translation model within the overall system architecture, nesting it within the training pipeline of the inference model. The translation model is integrated into the existing neural network structure, allowing the generation of additional learning data without requiring completely separate systems, thus managing complexity through hierarchical integration.

Inventive Principle:
Principle #7Nested doll (Nesting)

Data Source

PatentEP3846084A1Learning device, learning method, and computer-readable medium
Publication Date: 2021.07.07 KK TOSHIBA
  • EP3846084A1 patent drawingFigure 1
  • EP3846084A1 patent drawingFigure 2
  • EP3846084A1 patent drawingFigure 3

AI summary

An according to an arrangement, a learning device (1) includes an inference unit (13), a translation unit (12), a first update unit (141), and a second update unit (142). The inference unit (13) performs an inference task by using a first neural network (101a, 101b). The translation unit (12) translates second domain data into first translated data by using a second neural network (102). The first update unit (141) updates parameters of the second neural network so that a distribution that represents a feature of the first translated data approaches a distribution that represents a feature of the first domain data. The second update unit (142) updates parameters of the first neural network (101a, 101b) based on a second inference result output when the first translated data is input into the first neural network, a ground truth label of the first translated data, the first inference result, and a ground truth label of the first domain data.