Self-Training GANs for Semi-Supervised Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current semi-supervised learning methods in generative adversarial networks (GANs) lack effective mechanisms to exploit the visual characteristics of generated images and improve classification performance, particularly in self-training scenarios where synthetic data is not adequately utilized.
Innovation Solution
A self-training method and system for GANs that uses unlabelled and generated datasets to estimate data distribution and classification, automatically augmenting the training dataset by assigning labels to generated samples based on discriminator outputs, thereby expanding the dataset and enhancing training stability and performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional semi-supervised learning methods are used with GANs, then the model can generate synthetic images, but the classification performance does not improve due to inadequate utilization of generated data
Solution Approach 1:
The patent implements a self-training mechanism where the discriminator's predictions on unlabelled data are fed back into the system. High-confidence predictions are added to the training set as pseudo-labelled data, creating a feedback loop that continuously improves classification performance while effectively utilizing generated synthetic images.
Solution Approach 2:
The system performs self-training by automatically generating synthetic data, predicting labels for unlabelled data, and retraining the model without external intervention. The GAN and discriminator serve themselves by generating training data and improving their own classification capabilities through the self-training loop.
2Quantity of substance
If only unlabelled data is used for self-training, then the training process is simple, but the volume of training data is insufficient to improve discriminator performance
Solution Approach 1:
The patent uses the generator to create copies of real data in the form of synthetic images. These generated samples are then used as additional training data, effectively multiplying the available training data volume without requiring additional real-world data collection.
Solution Approach 2:
The generated synthetic data serves multiple functions: it augments the training set volume, provides additional examples for the discriminator to learn from, and enables the system to train on diverse data samples without requiring proportional increases in real data collection efforts.
3Productivity
If adversarial training and self-training are combined, then evolutionary learning is achieved, but the training process becomes more complex
Solution Approach 1:
The patent merges adversarial training and self-training into a unified framework. The generator and discriminator continue their adversarial game while simultaneously implementing self-training on unlabelled data, combining both training objectives into a single evolutionary learning process that improves both data generation and classification capabilities.
Data Source
AI summary
A method and system for augmenting a training dataset for a generative adversarial network (GAN). The training dataset includes labelled data samples and unlabelled data samples. The method includes: receiving generated samples generated using a first neural network of the GAN and the unlabelled samples of training dataset; determining a decision value for a sample from a decision function, wherein the sample is a generated sample of the generated samples or an unlabelled sample of the unlabelled samples of the training dataset; comparing the decision value to a threshold; in response to determining that the decision value exceeds the threshold: predicting a label for a sample; assigning the label to the sample; and augmenting the training dataset to include the sample with the assigned label as a labelled sample.