An End-to-End Multi-Modal Data Augmentation Framework Using Evolutionary Learning

An End-to-End Multi-Modal Data Augmentation Framework Using Evolutionary Learning

1 Mr. Sanskar H. Goenka, 2 Mr. Mahesh Y. Gondekar

Abstract: We present a novel, integrated framework for dynamic, multi-modal data ingestion, augmentation, and quality assurance using advanced generative adversarial networks (GANs) enhanced with evolutionary learning. Our platform leverages Deep Belief Networks (DBNs) for image feature extraction and transformer-based models for text feature extraction, fusing these modalities into a joint representation to drive a WGAN-GP with self-attention. An adaptive hyperparameter update rule based on a composite quality metric facilitates continuous improvement of synthetic data quality. Extended experimental results, a theoretical analysis, case studies, and discussions on limitations are provided. Critical implementation details have been abstracted to preserve the novelty while allowing future enhancements.