GAN Discriminator Predicting Continuous Labels

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing generative adversarial networks (GANs) lack the ability to effectively predict continuous labels for generated synthetic samples, particularly in semi-supervised regression tasks, which is inefficient and complex due to the need for converting continuous labels to discrete classes.

Innovation Solution

A GAN architecture is developed that includes a generator and a discriminator with a deep convolutional neural network (CNN) configuration, employing a feature matching loss technique and a combination of unsupervised and supervised loss functions to generate realistic synthetic samples and predict continuous labels effectively.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If classification techniques are used for semi-supervised regression tasks, then discrete class labels can be assigned to generated samples, but discretization errors are introduced and continuous label information is lost

Engineering Contradiction:
Improvelabel prediction accuracyVSAvoidmethod complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent uses the discriminator network from the GAN framework to copy its classification capability for predicting continuous labels on generated samples. The discriminator, already trained to distinguish real from fake samples, is repurposed to predict regression targets, eliminating the need for separate classification models and avoiding discretization errors.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The discriminator network serves multiple functions: it acts as both the discriminative model in the GAN minimax game and as a predictor for continuous regression labels. This multi-functionality reduces overall system complexity while maintaining high prediction accuracy for semi-supervised regression tasks.

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

2Quantity of substance

If GANs are used to generate synthetic training samples, then data scarcity is addressed, but the ability to predict continuous labels for generated samples is lacking

Engineering Contradiction:
Improvetraining sample quantityVSAvoidcontinuous label prediction capability
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent merges the GAN generation capability with label prediction by integrating the discriminator network that can simultaneously evaluate sample authenticity and predict continuous regression targets. This combination allows generated samples to inherit both synthetic data benefits and accurate continuous label assignments.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The discriminator provides feedback signals during GAN training that enable it to learn continuous label predictions alongside sample discrimination. The supervised loss from limited labeled data and unsupervised loss from unlabeled data work together to refine the discriminator's label prediction capability for generated samples.

Inventive Principle:
Principle #23Feedback

3Ease of manufacture

If supervised learning is used with limited labeled data, then model training is straightforward, but performance is constrained by data scarcity

Engineering Contradiction:
Improvetraining simplicityVSAvoidmodel performance
Core Design Contradiction:
Ease of manufactureVSReliability

Solution Approach 1:

The patent introduces the GAN generator as an intermediary that creates synthetic training samples from unlabeled data. These generated samples act as a bridge, allowing the model to learn from abundant unlabeled data while being guided by limited labeled data through the discriminator's supervised learning component.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The GAN framework performs preliminary action by generating synthetic training samples before the final supervised learning stage. This pre-generation of diverse training data expands the effective training set size, allowing subsequent supervised training to achieve better performance with limited actual labeled examples.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3602392B1Semi-supervised regression with generative adversarial networks
Publication Date: 2025.02.19 HUAWEI TECH CO LTD
  • EP3602392B1 patent drawingFigure 1
  • EP3602392B1 patent drawingFigure 2
  • EP3602392B1 patent drawingFigure 3

AI summary

Method and system for performing semi-supervised regression with a generative adversarial network (GAN) that includes a generator comprising a first neural network and a discriminator comprising a second neural network, comprising: outputting, from the first neural network, generated samples derived from a random noise vector; inputting, to the second neural network, the generated samples, a plurality of labelled training samples, and a plurality of unlabelled training samples; and outputting, from the second neural network, a predicted continuous label for each of a plurality of the generated samples and unlabelled samples.