Synthetic Environments for Artificial Intelligence (AI) and Machine Learning (ML) in Multi-Domain Operations
Abstract
The use of artificial intelligence solutions for Army field applications will rely heavily on machine learning (ML) algorithms. Current ML algorithms need large amounts of mission-relevant training data to enable them to perform well in tasks such as object and activity recognition and high-level decision-making. Battlefield data sources can be heterogeneous, encompassing multiple sensing modalities. Present open-source data sets for training ML approaches provide inadequate representation of scenes and situations of interest to the Army, in terms of both content and sensing modalities. There is a push to use synthetic data to make up for the paucity of real-world training data relevant to military multi-domain operations of the future. However, there are no systematic approaches for synthetic generation of data that provide any degree of assurance of improved real-world performance of the ML techniques trained on such data. The problem of effective synthetic data generation for ML raises deeper questions than that of artificially generating speech or imagery that humans find realistic. An Army Science Planning and Strategy Meeting held in December 2020 explored multiple technical issues in depth related to synthetic data generation and application to Army problems of interest.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 2021
- Accession Number
- AD1135395
Entities
People
- Celso De Melo
- Hamid Krim
- Raghuveer Rao
Organizations
- United States Army