The Emergence of Higher Scales in Complex Systems: The identification, growth, and training of informative macroscales in biological networks, artificial neural networks, and beyond
Abstract
Complex networks often have multiple valid scales of descriptions. This research provides formal tools for how to dimensionally reduce complex networks in ways that reveals their higher scales. Specifically, it focuses on how the uncertainty, or noise, can be minimized at the higher scales of networks and how this makes for more tractable and informative models. Networks with multiple scales are more robust, less noisy, and are more evolvable than networks lacking higher scales. Additionally, these techniques are applied to artificial neural networks, analyzing how higher scales develop during deep learning and how this provides benefits to generalizability and compression.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 09, 2020
- Source ID
- W911NF2010243
Entities
People
- Erik Hoel
Organizations
- Army Contracting Command
- Tufts University
- United States Army