Among the large amounts of data available to security analysts, there may be clues to an adversary’s attempt to acquire technologies for weapons of mass destruction. But there aren’t enough analysts in the world to sort through this information glut. Machine learning (ML) and artificial intelligence (AI) could automate the task of sifting through innumerable images, videos, sound, and text files, looking for needle-sized significant artifacts in haystacks of files billions of bytes large. For example, nuclear nonproliferation analysts could offload some of their work to systems trained to recognize the signatures of nuclear weapons technologies.
Livermore’s Barry Chen came directly to the Laboratory after receiving his Ph.D. in electrical engineering and computer science from the University of California, Berkeley in 2005. There he worked with automatic speech recognition, which depends on ML and AI technologies. “I wanted to find new ways to use these tools in the world,” he says. “Applying them to science and national security applications is very attractive and personally gratifying.”
Right away, Chen experienced the wide variety of work at Livermore, and the cross-disciplinary teamwork required to solve the complex challenges the Laboratory takes on. His first project was developing Computer Vision software to help find defects in each of the hundreds of laser optics used at the National Ignition Facility. He designed algorithms to detect distinctive bullseye patterns in images of NIF optics that are indicative of defect sites in upstream laser optics. He worked with other engineers and physicists to help keep NIF’s laser shots running without disruption.
Since then, he has spent a considerable amount of time applying ML to security applications, where the complexity and heterogeneity of data types pose problems to traditional ML approaches. He led the development of SparkPlug, a software toolkit for statistical density estimation of large sets of complex data. His team applied SparkPlug to model normal computer network behaviors for detecting anomalies in computer security applications.
Lately, Chen has been focusing on developing neural networks to model multimodal data—teaching them to map images, text, and video of similar concepts to proximal locations in a common vector space. With this capability, a nonproliferation analyst could specify a search for anything related to, for example, “centrifuge,” and the system can then scan massive collections of data, looking for technologies relevant to nuclear centrifuge development. Chen explains: “The challenges posed by these application areas are different from those faced by commercial interests. We are making a difference to both basic science and national security.”
Applies machine learning and artificial intelligence to security applications