Lawrence Livermore National Laboratory



Bruce Hendrickson

Bruce Hendrickson

Associate Director for Computation

Machine Learning and Laboratory Science Drive One Another

LAWRENCE Livermore is a data-rich environment. As the demand for sophisticated methods of analyzing and interpreting data grows, so too does the need to push the boundaries of data science—a field that goes beyond merely crunching numbers to exploiting sophisticated technologies such as machine learning (ML) to analyze data. Organizations all across the Laboratory apply data science techniques to scientific questions while also strengthening the field’s methodologies. The feature article, Machine Learning on a Mission, describes this dual pursuit in ML, in which computer algorithms learn from data to identify patterns, make inferences, and predict outcomes.

ML is a rapidly growing specialty, particularly in nearby Silicon Valley, California, where consumer needs propel technological developments. Around the world, major ML-focused conferences receive thousands of paper submissions, and the pace of scientific publishing is staggering. At the Laboratory, we contribute to this progress because ML has important implications for scientific data analysis and for our national security missions. Although many companies have pioneered ML methods for commercial use, the Laboratory’s problems often have different characteristics and demand unique solutions. Data-driven scientific decisions rely on researchers to create nuanced ML algorithms that can derive meaning from simulations, images, text, speech, video, and other data types. The article spotlights several ways in which Laboratory scientists adapt ML techniques to tackle application-specific challenges.

This adaptation can be incredibly difficult. For example, consider the ML task of classifying objects in an image, in which the computer tries to do what human eyes and brains can do easily. Indeed, the technology behind the ML algorithms that can successfully identify a cat, for instance, in a series of photographs is impressively complex. The computer must learn to distinguish feline-specific features from other pictured objects, the background, or interference. The process may entail pixel-level comparisons, color analysis, recognition of edges and boundaries, and more.

Now imagine programming a computer to detect the slightest three-dimensional anomalies in a two-dimensional scan of a vehicle at a border crossing. The computer must predict locations in the single image to search for unusual but nonspecifically defined objects that are not directly observable by humans. Compared to the cat-identification model, this cargo-scanning problem has fewer data to evaluate with more uncertainty and far greater consequences of error—such as incorrect or missing identification. Livermore is at the forefront of ML research that addresses real-world scenarios whose requirements go far beyond those satisfied by out-of-the-box ML tools.

Equally important is the advancement of ML as a science. Throughout history, many technologies have been developed and used before their mathematical and physical underpinnings had matured. Understanding the underlying principles of ML methods is important for reducing flaws, improving results, and enhancing our scientific knowledge. At Livermore, we strive to better comprehend how these methods work—such as how a specific prediction is made—so that we can have greater trust in them, and they can provide greater explanatory power. We seek assurance that ML models correctly extrapolate information and accurately reach conclusions. Unlike a movie recommendation, the decisions we make have significant consequences, and so we must have high confidence in them.

Fortunately, the Laboratory’s culture of multidisciplinary teamwork helps maximize the potential of this new technology. Our computer scientists, data analysis experts, and domain scientists work alongside each other to a degree unmatched in industry or academia. This collaboration across traditional disciplinary barriers, combined with powerful supercomputing capabilities, enables us to drive and respond to evolving technologies such as ML.

No doubt the ML landscape will look very different over the next few years, and it will become ever more important to the Laboratory’s missions in ways not yet foreseen. Given the increasing amounts of data being generated by experiments, simulations, and other sources, our researchers will continue to embrace and invent new data science and ML methods. This transformation of data-driven science is just beginning.