Smart algorithms for exploiting mass data will be available this year
- By Katherine Owens
- Aug 23, 2017
The first smart algorithms, compact bits of code embedded in weapons systems and sensor processors, will be unveiled later this year, said Col. Drew Cukor, Project Manager of the Algorithmic Warfare Cross-Functional Team, at the 2017 Defense One Tech Summit.
“Algorithms will be an important element of our weapons systems…these living, breathing algorithms, these small little bits of code that are going to fit inside of a larger ecosystem, are going to be sort of the brain,” explained Col. Cukor. “A constant learning environment [will] be created between data preparation, machine learning inside of significant computational power, and then finally… updating these algorithms,” he said.
The AWCFT, also known as Project Maven, is a fairly new initiative, first announced in April. Its purpose is to explore algorithms as a way to turn the influx of input data into manageable intelligence, according to a memorandum from the Deputy Secretary of Defense.
In the immediate future, the AWCFT program is working to finalize a total of 38 classes of object for algorithms to learn and recognize.
The 38 classes “represent the kinds of objects we need to detect in the counter-insurgency fight that we are involved in now… essentially all of the features that exist in society, because these are the weapons of insurgents,” said Col. Cukor.
The coding also goes beyond just image recognition, and algorithms are now being designed to engage in logical expression and scene recognition.
Before these algorithms can be fielded however, the data must be verified, or triaged, and labeled for the algorithm to “understand” what it is processing.
“These algorithms need large data sets and we are just starting labeling,” explained Col. Cukor. “We triage lots of imagery, and then it’s just a matter of how big our labelled data sets can get. We expect to see some phenomenal results just in this calendar year,” he reported.
Algorithms and artificial intelligence go hand-in-hand. When an algorithm is presented with labeled data and recognizes and categorizes it in one of the 38 object classes, it is performing a type of machine learning. However, algorithmic warfare does not mean replacing human decision-making with AI.
“As we implement what is probably the most primitive deep learning capability… there’s not going to be a selection of targets,” said Col. Cukor. “There’s going to be an advisory role, there’s going to be a collaboration role. There’s going to be detections and alerts. But the human is still a fundamental part of this change.”
The first algorithms that will emerge this year are designed to be compact, just about 75 lines of code as opposed to the thousands of lines of coding that make up the core functions of weapons systems.
For this reason, Col. Cukor does not believe that smart, compact algorithms will displace the existing complex software infrastructures. Nor will small software start-ups replace the core defense industry vendors.
The overall coding architecture of weapons systems that have been developed and maintained by defense contractors are still necessary for the function of the systems themselves.
Katherine Owens is a freelance reporter for Defense Systems