Progress in artificial intelligence (AI) is centered on data, of which there is paradoxically both too much and too little.
The technology L3Harris helped create to capture and transmit satellite imagery is moving faster than the U.S. military’s capacity to organize it. Machine learning algorithms provide a way to quickly identify objects in vast amounts of data but require human analysts to provide a contextual framework to learn from. By mastering this rigorous process, the military is poised to overcome the greatest challenge to creating a sophisticated AI platform.
For an AI solution to work, it requires three components: training data, neural network machine learning algorithm(s) and an integration platform. L3Harris has been entrusted by the U.S. Air Force (USAF) to develop the front-end workflows for generating and managing large data sets required to train USAF’s algorithms. USAF has also entrusted L3Harris' support for the back-end integration platform to interface with existing Department of Defense (DoD) systems.
Analysts now have the harrowing task of finding needles in a vast digital haystack of constantly streaming satellite imagery. There is an urgent need to automate this process. While neural network-based algorithms are commercially available to help, they require specialized expertise in the DoD and intelligence community data types and systems to be most effective.
Creating large enough data sets to effectively train algorithms is a resource-intensive endeavor. While the sheer quantity of training data needed is a challenge, of even greater importance is ensuring sufficient variation within the training data to represent real-world conditions. Synthetic data is disruptive technology that holds the key to breaking through the training data bottleneck for AI/ML solutions. By using computer-generated imagery that closely matches real-world imagery, training datasets with robust variations are rapidly generated. Synthetic data provides a scalable method of generating training data without having to devote extensive resources to manually harvesting, labeling and curating training data from real imagery. This value-add technology reduces costs, risks, biases, and enables a more accurate, trusted AI.
L3Harris’s expertise in high-fidelity, physics-based, radiometrically correct modeling and simulation of geospatial data makes them particularly qualified to create synthetic images that are indistinguishable from real-world imagery.
“We’ve been able to pivot and augment technology originally developed for sensor design and performance modeling, and apply it as a source of synthetic training data for AI. We are unique in that we do not simply make fake imagery that looks realistic to the human eye, rather we have tuned our synthetic data generation process specifically for neural network performance,” said Will Rorrer, Business Development Lead, L3Harris.
L3Harris’ capabilities will enable DoD AI practitioners to generate their own synthetic training data on demand through a web interface. The synthetic data will be used to train USAF’s algorithms to identify threats and hard-to-find objects. The trained algorithm is then hosted by the L3Harris platform to automatically analyze real collected imagery and evaluate performance.
“By integrating our synthetic data generation capabilities into an intuitive web-based interface, we enable AI developers to rapidly generate proven training data without needing an advanced understanding of image science," said Rorrer.
With precise synthetic data, L3Harris will fill USAF’s critical demand for advanced algorithm development and empower an army of digital analysts. By enhancing USAF’s machine learning capabilities, warfighters will have access to more accurate and actionable intelligence than ever before.
To learn more, please explore L3Harris' Geospatial Solutions below: