RPI ID: 2023-002-301, 2023-002-401
Innovation Summary: This technology introduces a novel embedding framework for representation learning that leverages high-order mixed statistical moments to capture complex data relationships. The system enhances feature extraction by encoding richer structural and distributional information from input data. It is designed to improve performance across various machine learning tasks, especially in scenarios with limited labeled data. The method is compatible with self-supervised and supervised learning architectures.
Challenges / Opportunities: Traditional embedding techniques often rely on first-order statistics, which may miss nuanced patterns in data. This invention addresses the challenge of capturing higher-order dependencies to improve model generalization. It presents an opportunity to advance representation learning in domains such as natural language processing, computer vision, and bioinformatics. The approach also supports scalable learning in data-rich but label-scarce environments.
Key Benefits / Advantages: ✔ Captures complex data relationships using high-order mixed moments ✔ Enhances feature richness and model accuracy ✔ Supports both supervised and self-supervised learning ✔ Reduces reliance on labeled datasets ✔ Scalable across diverse data modalities
Applications: • Representation learning in AI systems for text, image, and biological data analysis
Keywords: Representation learning, high-order moments, embedding, self-supervised learning, feature extraction, machine learning
Intellectual Property: WO2024015625A1 (Application PCT/US2023/027876), Published July 17, 2023; Application 18/994694 Pending