Pretty fly for an AI: Bioboffins use machine learning to decipher fruit flies' brains

Code cracks neuron riddle

Scientists in the US have developed a computer program called JAABA that uses machine learning to map groups of neurons responsible for the different behaviors observed in tiny fruit flies.

The brain is a tangled mess of neurons that continues to mystify neurologists. Creating a detailed map allows scientists to learn the anatomy of the brain in the hopes that they can understand how behavior manifests neurologically.

Drosophila melanogaster, or fruit flies, are a good place to start. The insects have a poppy seed-sized brain with 100,000 neurons compared to the 100 billion in human brains.

The only concerns in the flies' lives are seeking out ripe fruits and vegetables and mating. Although their behavior is simplistic, understanding it is still a complex job. Researchers from the Howard Hughes Medical Institute, a non-profit medical research laboratory in Ashburn, Virginia, decided it was a task ripe for machine learning. Details have been published in the journal Cell.

Previous work had identified what neurons were associated with particular areas of the flies' brains. The whole brain was split into different groups of neurons, which were activated to make the flies behave a certain way.

Jonathan Hirokawa, co-author of the paper and previously a lab technician at the Howard Hughes Medical Institute, said: "Sometimes you'd get flies that would all turn in circles, or all follow one another like they were in a conga line."

When other neurons were fired, the flies would huddle close together, avoid each other or chase each other around. The behavior of 400,000 flies was filmed, producing 20,000 videos. The computer program, Janelia Automatic Animal Behavior Annotator or JAABA, tracked the movements of each fly.

The behaviors were manually labelled. Over 500TB of video data was used to train classifier algorithms to associate a certain type of motion to a recognized behavior. Since the researchers know what areas of the brain have been stimulated, they can create detailed brain maps to find out what neurons are responsible for specific behaviors.

How the brain maps were produced (Image credit: Robie et al)

The researchers report a new discovery identifying the nerve cells linked to hostile responses between females.

"There have been some reports of female aggression, but not females chasing other flies," said Alice Robie, co-author of the paper.

Using JAABA has generated thousands of results pinpointing what neurons are associated with a suite of 14 behaviors, including jumping, walking, wing flicking and wing grooming.

Machine learning excels at learning the relationships between different variables and pattern matching. It is estimated that it would take humans some 3,800 years to analyze the 225 days of video footage – something that JAABA can do in hours.

"We processed the videos on a cluster of machines during the night after collecting them during the day. The total compute time was probably on the order of 20 years, but we could run it on my computers in parallel automatically," Kristin Branson, co-author of the paper, told The Register.

For now this technique is only applicable for simple animals like fruit flies, Branson explained.

"These methods are based on genetic manipulations which can only really be done in model organisms like fruit flies. However, we are hopeful that the general principles of neural computation that can be discovered by studying simpler organisms will be shared with humans."

Robie said the new results show the benefits of mixing AI with science. "This is what happens when you put biologists and computer scientists together." ®


Biting the hand that feeds IT © 1998–2017