Applying artificial intelligence to an old tradition
“Ah-ah-ah-CHOO!”
With spring comes waves of wind-blown pollen and spores, much to the chagrin of hay fever sufferers across the country. This year marks 101 years since the origination of palynology, the discipline that examines these allergy-causing culprits.
Quantitative palynology arose in 1916 when Lennart Von Post, a Swedish palynologist, published a scientific journal article on fossilized pollen preserved in peat bogs.
“Von Post was the first to suggest that you could go from pollen to a community reconstruction,” said Dr. Surangi Punyasena, a palynologist at the University of Illinois at Urbana-Champaign.
Palynology has come a long way since then. Its data has proved useful in many different fields, including allergy research, ancient environment reconstruction, climate change studies, and forensic science, to name a few.
Now, this old tradition is meeting modern technology, as palynologists begin to harness artificial intelligence to assist with their work.
The identification of pollen and spores is an incredibly time-intensive task. These reservoirs of plant reproductive information are mostly distinguished by their shape and surface texture—features too small to scrutinize with the naked eye. Consequently, palynologists must spend countless hours behind a microscope to both learn and identify pollen and spore types.
“The actual skill of identifying is a visual skill. I see it as a bit of artistry in the sense that you have to have a very strong visual memory of hundreds of types,” Punyasena said. “It takes years, almost even a lifetime to recognize and learn these types,” she continued.
Unfortunately, that identification expertise is lost with each generation, as palynologists retire and pass away.
“In some ways, we re-discover old knowledge over and over again, and we probably don't need to if we were able to more efficiently pass that knowledge on,” explained Punyasena.
Recent work suggests that artificial intelligence, specifically neural networks, may hold the answer.
“Computer scientists are not just working on a computer vision, but are working on this whole area of knowledge acquisition,” said Punyasena.
Punyasena believes that neural networks—computer systems created to mimic the human brain—could serve as an expertise intermediary between generations of palynologists.
“Neural nets are becoming increasingly popular in a lot of types of image analyses. They come pre-trained. They are already efficient when you start working with them and they just become more efficient at the process as you train them,” Punyasena said.
Punyasena and her research team train neural nets to recognize, isolate, and identify pollen and spores present in high-resolution, stacked images. To do so, they identify and tag features on images within the system.
“You can essentially record every decision you're making, and so as you learn the system learns, and the whole system becomes smarter,” Punyasena said.
As more information is added to the neural net’s library, its ability to recognize and identify pollen and spores within images grows. By using these progressively “smarter” neural nets, new generations of palynologists would be trained in the most current knowledge of the discipline.
Palynology neural nets would also provide a big boost for expertise exchange across the entire discipline.
“There's a push in science more generally towards this collective sharing of information,” Punyasena said. “To more efficiently transfer that knowledge from palynologist to palynologist, from lab to lab, from generation to generation,” said Punyasena.
Still, it will take substantial time and effort to build up these neural nets.
“Right now, what's limiting us is the quality of the image we can take, and the speed at which we can take that image. And the other thing that's limiting us is that we have a relatively small number of people working tagging and treating these images,” Punyasena explained.
Once the foundation has been laid, however, these neural nets could open new areas of inquiry that were not feasible before.
“It's not just to make work flow quicker. It's to expand the kind of questions we can ask,” said Punyasena. “We could actually work with more samples and on a scale that most paleoecologists don't work with in terms of numbers and breadth of time and geography.”
Editor’s notes:
To reach Surangi Punyasena, call 217-244-8049; email spunya1@illinois.edu.
