On Data-Driven Saak Transform: Theory and Applications
Abstract
This US Army Combat Capabilities Development Command Army Research Laboratory's External Collaboration Initiative is a joint effort with the University of Southern California. Combining the team members experience in machine learning, signal analysis, computer vision, and perceptual communication and computation led to insights on machine-learning mechanism and the impetus to develop an innovative theory and mathematical framework, the Subspace Approximation with Augmented Kernels (Saak) transform for deep neural-network architecture. The Saak transform is an entirely new signal theory based on insightful interpretations of the deep-learning mechanism. We conducted the fundamental research on Saak transform theory and its two important extensions: Subspace Approximation with Adjusted Bias transform and Successive Subspace Learning. We unified them under a common framework of interpretable subspace learning. We applied developed theories to several militarily relevant tasks and scenarios including image classification, recognition, defense against adversarial attacks, and biometric facial-data processing.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 08, 2021
- Accession Number
- AD1128133
Entities
People
- Suya You
Organizations
- United States Army Research Laboratory