Understanding/Applying Neural Models to New Domains
Abstract
This Final Performance Report summarizes the work accomplished and lessons learned during the original and subsequent engineering change of this agreement. Specifically for the original agreement, this effort proposed three tasks: (1) understand the External Plexiform Layer (EPL) model in greater detail; (2) train the EPL model; and (3) apply the EPL model to image recognition / object identification. The engineering change (1) further investigated various training approaches / architectures to generate hyperdimensional vectors (> 1,000 bits) as the trained output instead of classification vectors; (2) applied these new models to additional datasets; and (3) investigated issues/concerns that needed to be addressed to transfer this technology to operational problems. We also investigated several newer models/architectures designed around lifelong learning and dealing with the ever-changing real world. These newer models include Random Networks; Kernet-- a Kernel-based model; Sparse High-Level-Exclusive, low Level-Shared (SHELS) and OpenCon -- Open-world Contrastive Learning. Detailed discussions, results and pros/cons of each model are provided.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 05, 2024
- Accession Number
- AD1222871
Entities
People
- John Salerno