Researcher builds AI into Danish military software
Can machine learning help reveal your adversary’s military strategy? Henrik Pedersen is working on answers to this NP-hard computer science problem. He is the researcher who is integrating AI into systems for military decision-making support.
Article published in Version2:
Henrik Pedersen is a part-time lecturer at the Department of Computer Science at Aarhus University, where he teaches his students in deep learning. He holds a PhD in AI and medical image processing and did his thesis in cardiac and brain scanning. As a team lead at the Alexandra Institute, one of seven Danish government-approved Research and Technology Organisations (or ‘GTS institutes’), he built an algorithm that could estimate the weight of a slaughter pig with a degree of accuracy of +/- 3%. It required 3.7 million images.
At Systematic, Henrik Pedersen is the AI researcher who is bringing AI-based innovation into the company’s digital solutions for NATO armed forces. The solutions include the advanced command-and-control (C2) system SitaWare, which helps officers maintain an overview under pressure and thus make the right decisions at the right time.
Military textbooks in the computer
At the moment, Henrik Pedersen is looking at whether, for example, AI can help officers identify an adversary’s strategy. Basically, it involves feeding all the Russian military textbooks into a computer and then comparing the strategies with what is happening on the battlefield. When the AI recognises specific patterns, it can possibly reveal what the overall plan is.
"Recognising a military doctrine with machine learning is what is known in computer science as an ‘NP-hard problem’ (non-deterministic polynomial-time) – i.e. a problem which you cannot find the optimum solution to without first having tried out all the possibilities. It is a difficult computer science problem. However, by using machine learning, we can arrive at what we call an ‘approximative solution," says Henrik Pedersen.
Henrik Pedersen says that he and other developers at Systematic are working with research-based innovation in areas which are basically unknown territory. Among other things, they have developed a system for image recognition which a SitaWare system user can train on their own.
Unique image search engine
An officer can, for example, feed a few aerial photos of destroyed tanks into the system and ask it to find similar photographs in a large intelligence database. When the system gets back with a selection of images, you select the correct ones and the false positives and then implement a search. After two or three rounds, the AI has trained itself so well that the error rate is reduced to an acceptable level.
"It started as an idea that popped into my head. So, I built a prototype which I showed to our product management team, and I’ve now integrated it as a feature in our SitaWare Insight system. Nothing like it exists elsewhere," says Henrik Pedersen.
Runs in a closed environment
ChatGPT, Bard and other large language models which can be accessed online are a no-go in a military context. Instead, Henrik Pedersen and his colleagues at Systematic build their own AIs based on open source models and typically within computer vision and natural language processing. The AIs run in a closed environment, and are trained to solve very specific tasks. Henrik always builds prototypes to check whether his ideas work in practice.
"When you code a prototype, you often discover something that you hadn’t previously thought of. Moreover, functional prototypes are extremely good at demonstrating your idea to colleagues and customers."
At the moment, Henrik Pedersen is the only full-time employee in Systematic Defence’s Product Lab. Soon, the line-up will double in size when he is joined by a new colleague who recently qualified with a degree in cognitive science.
"Things are developing at such a pace within machine learning that it’s difficult to keep up. It’s good if more than one person is assigned to the task. Especially when we’re also coding prototypes. That’s the most important thing of all: It’s only once you’ve finished building it that you know whether or not it works."
