Researchers use moss to overcome machine-learning limitations
Deep-learning algorithms can identify objects and faces better than humans in some cases but are limited in their ability to recognize amorphous forms, like grasses and trees, that can take different shapes and sizes and are continually changing. But a team, led by Takeshi Ise at the Kyoto University in Japan, has developed a new technique that will help machines overcome these limitations, per the MIT Technology Review. The method centers on teaching machines to recognize different types of moss, a plant that doesn't take a well-defined, distinctive shape.
Why it matters: The method could be used to better recognize trees and crops in aerial photos, which would be valuable for monitoring agriculture and for conservation and land management efforts.
How they did it: Ise and his team started by photographing three different types of moss, both individually and in places where they were grouped with other plants. They then worked to create an algorithm that could distinguish the different types of moss by meticulously labeling the different versions and feeding it into the deep-learning machine.
The results: The algorithm quickly learned to identify each type of moss when grouped among others in a single image. "The model correctly classified test images with accuracy more than 90%," the team said.
Limitations of study: The algorithm recognized some types of moss better than others. For example, the types of moss that are more amorphous with less defined forms of growth were harder to identify than those that had a more distinctive shape. But the team said they plan to continue improving the accuracy of their method by experimenting with how the photos are taken.