GAN Training with Shared Weights and Constraint Functions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Generative adversarial networks (GANs) struggle to generate images that accurately represent underlying distributions when training data is under-represented or lacks diversity, leading to difficulties in learning to generate representations for categories without sufficient training samples.
Innovation Solution
The method involves training a GAN with two generators and a discriminator, where one generator learns from available categories and shares neural network weightings with another generator to learn from unavailable categories using a constraint function and knowledge loss, enhancing the diversity of the training distribution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If GAN is trained with limited and under-represented training data, then data collection cost is reduced, but the generated images fail to represent the underlying distribution accurately
Solution Approach 1:
The patent introduces a constraint function as an intermediary that encodes domain knowledge about the underlying distribution. This constraint function acts as a mediator between the limited training data and the generation process, guiding the generator to produce images that reflect the true distribution even when training data is insufficient or biased.
Solution Approach 2:
The patent implements a feedback mechanism where the constraint function continuously provides guidance to the generator during training. The generator receives feedback from the constraint function about deviations from the underlying distribution, allowing it to adjust its outputs to better represent the true distribution despite limitations in training data.
2Ease of manufacture
If GAN is trained without category-specific training samples, then data collection effort is reduced, but the generator cannot learn to generate representations for unavailable categories
Solution Approach 1:
The patent makes the constraint function universal by encoding general domain knowledge that applies across multiple categories. This single constraint function can guide the generator to produce images for both available and unavailable categories, eliminating the need for category-specific training data while maintaining generation capability across the entire distribution.
Solution Approach 2:
The constraint function serves as a mediator that bridges the gap between limited training data and the need to generate images for unavailable categories. It provides the necessary guidance and structural information that enables the generator to create plausible images for categories not present in the training set.
3Device complexity
If GAN uses shared neural network weightings between generators, then model complexity is reduced, but the ability to learn from different categories simultaneously is challenged
Solution Approach 1:
The patent merges the neural network weightings of multiple generators into a single shared set of weights. This consolidation reduces model complexity and computational resources required while the constraint function ensures that the shared weights can effectively learn from and generate images across different categories without interference.
Data Source
AI summary
The disclosure provides a method for training generative adversarial network (GAN), a method for generating images by using GAN, and a computer readable storage medium. The method may train the first generator of the GAN with available training samples belonging to the first type category and share the knowledge learnt by the first generator to the second generator. Accordingly, the second generator may learn to generate (fake) images belonging to the second type category even if there are no available training data during training the second generator.