Variable Mix Ratio Data Generation for GAN Convergence

The patent applies preliminary action by pre-processing the training data through mixing real and fake data at different ratios before feeding it to the Discriminator. The mix data generation unit creates augmented training samples in advance, allowing the Discriminator to learn from diverse data distributions without requiring extensive training time. This pre-prepared mixed data serves as enhanced training material that accelerates the learning process while maintaining high discrimination accuracy.

The patent implements local quality by varying the mix ratio according to the position of each data element within the data sequence. Instead of applying a uniform mix ratio to all data elements, the system adjusts the mix ratio locally based on positional information, allowing different regions of the training data to have optimized mixing characteristics. This position-dependent mix ratio strategy enhances learning efficiency by providing spatially varied training examples while maintaining manageable processing complexity through a systematic approach.