More Synthetic Data Needed to Improve AI.

• Author: Tadjdeh, Yasmin
• Position: Algorithmic Warfare

* Developers of artificial intelligence platforms have promised that such systems will revolutionize the future of warfare, but experts and analysts said that more synthetic data is needed for that promise to be realized.

Pedro Rodriguez, a senior research scientist at the Johns Hopkins University Applied Physics Laboratory, said artificial intelligence and machine learning are being used extensively to analyze full-motion video, but the quality of the data is not as high as many would expect. Oftentimes they consist of low-resolution images and obstructions such as clouds.

"There is clearly a role for simulated data to be played in this environment," he said during a panel discussion at a defense and security conference hosted by the Association of Unmanned Vehicle Systems International in National Harbor, Maryland.

Gregory Allen, an adjunct fellow at the Center for a New American Security's Technology and National Security Program, said artificial intelligence platforms will need large amounts of synthetic data in order to improve their functionality.

"The best machine learning and AI algorithms that we have today--the ones with the most amazing super human performance--are all incredibly data hungry," he said. "They require large data sets upon which to train the algorithm in order to have a high-performance system that you can deliver."

Artificial intelligence systems can only function based on what they know and have been exposed to, he noted.

"If you train your AI system to analyze drone video data at one altitude and one level of cloudiness, then it is going to have very high performance at that altitude and that level of cloudiness," he said. "But if that happens to be different from the operational environment you encounter on the field, then the performance is going to fall off a cliff."

There is a need to create synthetic information that can provide a system with a variety of environments it will be expected to encounter in the field, he said.

An artificial intelligence system can generate its own synthetic data, Allen noted.

For example, AlphaGo--an AI system developed by Google DeepMind--was able to beat the world's Go champion, he said. Go is a two-player board game.

"A large portion of the data set upon which that system was trained was generated through self-play," he said. "The system was playing the Go game against itself and then feeding that into the total library data corpus of games upon which it was developing its AI...