What you should know about the future of machines

By Mark Wisinger, senior analyst

2017 saw machine learning become the de-facto in-vogue technology, whether the conversation was about data, cybersecurity or even traditional business systems.

In December, Google’s AlphaZero chess engine, utilizing Google’s DeepMind AI, crushed the incumbent chess engine champion, Stockfish. Google’s DeepMind relies heavily on machine learning – the AlphaZero chess engine did not start with any human knowledge, yet was able to learn how to beat Stockfish in 400 hours through machine learning. It’s a clear victory for machine learning – but one that’s easy to simulate. This is a much easier use case than identifying noise from cyber threats or prioritizing and cleaning multiple forms of data.

At immixGroup’s Government IT Sales Summit, we hosted a discussion on artificial intelligence (AI) and machine learning with Ron Gula, president and co-founder of Gula Tech Adventures and former CEO and co-founder of Tenable Networks, Dr. William Vanderlinde, chief scientist at the Intelligence Advanced Research Projects Activity (IARPA) and Rich Friedrich, senior director of cyber security analytics at Micro Focus Government Solutions.

Here are key takeaways to keep in mind when you discuss machine learning and AI with your government customers:

Machine learning is young and growing

Gula pointed out that the technology is not mature enough for cyber prediction. What’s marketed as machine learning is actually advanced reporting or crowdsourced sharing.

Friedrich warned of how much snake oil is often found in the IT industry. Machine learning is bandied about in marketing materials, but it’s not quite there in production and real-world environments.

Nevertheless, machine learning is a focus for IARPA and the Defense Advanced Research Projects Agency (DARPA)—the research arms of the Department of Defense and the Intelligence Community. Dr. Vanderlinde said 60 percent of IARPA’s research programs involve machine learning in one way or another, helping to fuel a strong push in the private and public sectors for these capabilities.

Where there are limitations, there are opportunities

There’s a lot of difficulties training systems and software to accurately identify meaningful data or pertinent threats. Tech companies are taking lots of data and metadata and dumping it into tools like the Elastic Stack to generate conclusions but this model just doesn’t work for real use cases yet, Gula said. A big limitation for training software comes down to the data – machine learning shines where you have large, labeled datasets but breaks down when the data is unstructured, labeled or limited in scope.

In the last 6 months, IARPA has hired someone focused on the issue of labeling datasets – it’s just that important for machine learning, said Dr. Vanderline. And he added that we’re going to need additional advances in data management and hygiene before machine learning really starts making headway in production environments.

Another limitation with machine learning and AI is that machines can have a tough time differentiating between correlation and causality–Just because people are using umbrellas when it rains doesn’t mean the umbrellas are causing the rain. This can really lead to false positives and it’s very tough to keep false positives down to an acceptable rate.

There’s no doubt machine learning is still in the early stages and has a long way to go. Tech companies should keep an eye on these limitations when speaking to government customers and industry partners. There should be more great advances in 2018 in this space, perhaps this time beyond the chess board.

Watch the entire machine learning and artificial intelligence session here.

Like what you read? Get more insight on selling technology to the government by subscribing to the Government Sales Insider blog.