Richland scientists are teaching computers to learn — and advancing scientific frontiers

PNNL researchers are exploring the use of deep learning in a range of scientific applications, including helping interpret signals from radioactive decay events by separating those of interest from background noise.
PNNL researchers are exploring the use of deep learning in a range of scientific applications, including helping interpret signals from radioactive decay events by separating those of interest from background noise.

When you watch young children learning to sort shapes or recognize letters, you can see how layers of knowledge build upon to create new understanding. Humans excel in this regard.

At the Department of Energy’s Pacific Northwest National Laboratory, we seek to endow computers with a similar capability by advancing a field of artificial intelligence called deep learning.

In deep learning, we combine computational power and a structured, automated way to build analytical models that mimic the brain’s ability to systematically incorporate layers of knowledge.

This involves developing computer programs that can learn from the data being analyzed, automatically identifying patterns and even making decisions — all with minimal human intervention.

Layers of knowledge

Here’s an oversimplified description of how a PNNL-developed tool called Sharkzor works.

First, a user provides the system with a few examples of images in various categories, such as animals and food, so that the tool’s algorithms can reliably recognize and sort new images into these user-defined categories.

As it sorts, the algorithms are continually and automatically updated and refined to better represent the features that the computer “sees” as it compares new images to those already labeled.

Ultimately, the system can accurately recognize and sort thousands of images into user-defined categories in just seconds.

With a little more training, it can distinguish finer details and sort images into increasingly narrow sub-categories, separating dogs from cats and even Border Collies from Australian Shepherds, for example.

012819 Ashby Flynn.jpg
PNNL earth scientist Donna Flynn collaborated with computational scientists to apply machine learning to identify clouds more accurately. Their model can identify clouds with nearly as much precision as an expert’s eye, in just a fraction of the time. Courtesy PNNL

Health, national security, environment

At PNNL, we are teaming data scientists with researchers in various domains to explore real-world applications of deep learning in many areas, including health, the environment and national security.

For example, in collaboration with Fred Hutchinson Cancer Center and the Seattle Cancer Care Alliance, PNNL researchers are exploring the use of deep learning to identify malignant tumors in breast cancer screens.

In a study of more than 5,000 patients that involved comparing radiologist reports, pathology reports and the outcomes captured in long-term cancer registries, their tool is nearing the 80 percent accuracy achieved by radiologists.

In another application, researchers seeking to improve climate models have used deep learning to identify clouds in LIDAR data with twice the precision of conventional techniques.

Scientists began by manually labeling data images pixel by pixel to “ground truth” the data and train the model.

They then ran a series of computations to identify the cloud boundaries. As the model learned, it compared its results against the hand-labeled data and adjusted accordingly — producing better results as more data was analyzed and compared.

On the national security front, researchers are using deep learning to analyze the energy pulses collected at PNNL’s Shallow Underground Laboratory to identify signals of radioactive decay that may indicate nuclear test activity versus those of routine events.

In multiple tests, the model accurately sorted the “good data” from background noise 99 to 100 percent of the time.

Pushing science boundaries

PNNL is exploring a variety of additional applications for deep learning, including real-time control strategies for the electricity grid, analyzing oil and gas pipeline sensor data to better prioritize preventive maintenance, and generating novel cyberattack patterns to design more effective defenses.

Researchers are developing similar approaches to identify chemical compounds in complex mixtures, to analyze geospatial sensor data to detect invasive grasses and even to scan social media to predict how a flu epidemic might spread.

Just as deep learning builds upon layers of knowledge to create new and better results, PNNL researchers are adding layers of knowledge and experience about this exciting field that will push the boundaries of science and engineering.

Steven Ashby is director of Pacific Northwest National Laboratory in Richland.