Understanding the Fragility of Neural Network Representations for MCM
Abstract
In this effort, PSU/ARL proposes to begin modeling the representation produced by deep neural networks (NNs) with the goal of understanding their fragility. NN algorithms are sensitive inthree ways:1. Input information quality (i.e. sensor fidelity, simulation fidelity, label quality, intersensormismatch in the case of transfer learning, etc)2. Model training approach (i.e. stopping criteria, optimization algorithm, etc)3. Model architecture (i.e. Alexnet, ResNet, Densenet, NASNet, Random, etc.)PSU/ARL wants to model the NN performance similar to work done in performance estimation, but with the explicit goals of being able to: characterize weaknesses in what the network models describe the weakness of the modeling in a practical manner mitigate the weaknesses as best as possible
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 17, 2020
- Source ID
- N000142012406
Entities
People
- Isaac D. Gerg
Organizations
- Office of Naval Research
- Pennsylvania State University
- United States Navy