Generative Adversarial Network for Synthetic Validation Data Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In applications with limited data, such as tracking tasks, generating validation data for recurrent neural networks is challenging due to the spatio-temporal nature of the data, making it time-consuming and costly to label and identify tracklets, which hampers the performance of these networks.
Innovation Solution
An evaluation device and method that generate artificial validation data by learning to classify real and artificial data sets indistinguishably, using a generative adversarial network to produce data that mimics the distribution of real data, thereby duplicating and expanding the available validation data for training recurrent neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling and identification of tracklets is used to generate validation data, then data accuracy is improved, but time consumption and cost increase significantly
Solution Approach 1:
The patent uses Generative Adversarial Networks to create synthetic copies of real tracklet data. The generator network produces artificial validation data that mimics the statistical properties and characteristics of manually labeled data, while the discriminator network evaluates the authenticity of generated samples. This copying approach eliminates the need for time-consuming manual labeling while maintaining data quality.
Solution Approach 2:
The system enables self-service data generation by automating the validation data creation process. Instead of requiring human annotators to manually label tracklets, the GAN framework allows the system to automatically generate synthetic validation data with proper labels through the interaction between generator and discriminator networks, significantly reducing time consumption and cost.
2Reliability
If more validation data is collected to improve network performance, then training quality is improved, but data availability remains limited in specialized applications
Solution Approach 1:
The patent addresses data scarcity by creating synthetic copies of limited real tracklet data using GANs. The generator network learns the underlying distribution of available validation data and produces additional synthetic samples that expand the training dataset. This allows specialized applications with limited real data to achieve better network performance through augmented data availability.
Solution Approach 2:
The system changes the parameters of the data distribution by learning from real tracklet characteristics and generating synthetic data with varied parameters. The GAN framework allows exploration of the data manifold by generating samples with different spatial and temporal parameters while maintaining the essential characteristics of genuine tracklets, thereby expanding data availability without requiring additional real-world collection.
3Quantity of substance
If synthetic data generation methods are used to expand data availability, then data quantity is improved, but distinguishing generated data from real data becomes challenging
Solution Approach 1:
The patent implements a feedback mechanism through the discriminator network that evaluates generated tracklet data and provides feedback to the generator. The discriminator attempts to distinguish real from synthetic data, and its feedback signals guide the generator to improve the realism of generated samples. This adversarial feedback loop ensures that synthetic data maintains high distinguishability quality while expanding overall data quantity.
Solution Approach 2:
The system employs dynamic interaction between the generator and discriminator networks that evolves over training iterations. As training progresses, the generator dynamically adapts to produce more realistic synthetic data, while the discriminator dynamically improves its discrimination capability. This dynamic adversarial process ensures that synthetic data remains indistinguishable from real data while continuously expanding the available dataset.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
Evaluation device (10) for generating validation data for an intelligent algorithm (2, RNN) comprising an input interface (11) for receiving first data (x) from at least one environmental sensing sensor (1), a memory (12, 21) comprising a program (13) comprising a first program section (G) for generating second data (G(z)) depending on the first data (x), and a second program section (D) programmed to receive the first data (x) and the second data (G(z)) as input in order to learn, during repeated execution of the program (13), a classification of the first data (x) as data from the at least one environmental sensing sensor (1) and of the second data (G(z)) as generated data, wherein the first program section (G) is programmed to learn, during repeated execution of the program, to generate the second data (G(z)) in such a way as tothat a second distribution function (p_G) of the second data (G(z)) resulting from this data generation is indistinguishable from a characteristic of the first distribution function (p(x)) in one characteristic of the second distribution function (p_G), wherein the evaluation device (10) is configured to execute the program (13), and an output interface (15) is provided to supply validation data for the intelligent algorithm (2, RNN) from the second distribution function (p_G), wherein the validation data includes data from the second distribution function (p_G) for training the intelligent algorithm (2, RNN). The invention also relates to a method for generating validation data and a computer program product that executes the method.