Cell News—Seeing the leaves for the forest

A machine learning system can find patterns in leaf venation in finer detail than any expert eye. ASCB photo by John Fleischman

A machine learning system can find patterns in leaf venation in finer detail than any expert eye. ASCB photo by John Fleischman

Did you have a high school biology teacher who was keen on taxonomy and keying out tree species? If you discovered that the real forest and the keyed forest in your handbook didn’t always match, you might enjoy an ingenious new paper in PNAS from a team of experts in geoscience, machine learning, ecology, paleobotany, and visual recognition led by Peter Wilf at Pennsylvania State University and Thomas Serre at Brown University’s Institute of Brain Science. The team devised a machine learning system that keys out angiosperms in ways that would have left Linnaeus in shock. The system sorted out scanned tree leaves, “clear specimens” (bleached of color and stained to reveal their intricate vein patterns), drawn from “vouchered” collections of 7,597 specimens from 2,001 genera and representing 6,949 species. The machine generated its own rule book by which it plotted key locations, compiled heat maps of critical differences, and then divided the specimens into evolutionary families and orders with an accuracy 13 times greater than chance. It was able to distinguish evolutionary clades that DNA data is currently unable to resolve.

 

The researchers say that the system could be of great use to evolutionary biologists trying to make connections between previously unrecognized branches and for paleobotanists dealing with single leaf fossils of early common ancestors. It also underscores the potential for machine learning systems to spot patterns in data in far finer detail than by any human expert.

About the Author: