Fast and Reliable Object Identification via Guaranteed Learning on Neural Networks
Abstract
Army networks and systems are generating and collecting vast amounts of data with different modalities, sampling rates, and accuracies continually. The processing and information extraction from these massive amounts of data for real-time decision making becomes exceptionally challenging. This project will develop fast learning methods from massive amounts of data for anomaly detection and object identification. One distinctive feature of this project is that it will establish the theoretical and algorithmic underpinning for deep learning so that the developed methods in this project can be applied to diverse systems with analytical performance guarantees. This project explores convolutional neural networks and graph convolutional neural networks for anomaly detection, object identification, and human prediction. It will also develop pruning and normalization methods to enhance the learning rate without sacrificing performance. This proposal provides systematic analyses of three intervened aspects of deep learning, algorithmic efficiency, generalization to test data, and sample complexity. The special features of the methods developed in this project include: 1. Guaranteed accuracy of the learned models on unseen testing data. 2. Fast training algorithms. 3. Theory-guided approaches for fast learning with limited data samples. This project will enhance the transparency of neural networks and provide algorithmic and theoretical tools to design learning methods and analyze the learning performance. It will resolve the concerns about the accountability and transparency of AI technologies and enable its reliable implementation on Army applications. Motivated by anomaly identification and prediction, this research is fundamental in nature, and the developed results apply to the general classification and prediction tasks in other domains.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 25, 2021
- Source ID
- W911NF2110255
Entities
People
- Meng Wang
Organizations
- Army Contracting Command
- Rensselaer Polytechnic Institute
- United States Army