Third Wave Deep Learning Methods For Physical RealisRc Data
Abstract
This seedling will directly address DARPAÕs Third Wave AI Initiative. BlueLightAI, Inc will build novel Deep Learning tools that contain contextual knowledge associated to a modality, incorporate prior knowledge through geometry, learn faster and more robustly, generalize better, be less susceptible to adversarial attack, give better compression of both data and algorithms thus allowing for sparser computation, will be ideally suited for edge computing, and will be equipped with diagnostics capabilities that permits analysis and provides transparency. This seedling addresses many of the current limitations of AI and Deep Learning. These include requiring huge amounts of training data, which is not well suited to edge computing and to small sample sizes, failing to generalize well across similar dataset types, lacking transparency and diagnostic tools, being vulnerable to adversarial attacks, addressing only a limited number of cognitive tasks, and where human/machine interactions are limited. Success will enable Deep Learning to improve human/machine interactions, generalization and robustness, transparency and diagnostics, operate better on smaller datasets, and support wider ranges of cognitive tasks. In particular, success will offer a paradigm shift from training on big data for small insights (those that generalize poorly) to training on small data for bigger insights (those that generalize better).
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 04, 2023
- Source ID
- W911NF2110315
Entities
People
- Gunnar E Carlsson
Organizations
- Army Contracting Command
- Defense Advanced Research Projects Agency