Researchers at Canadian Institute for Advanced Research developed a neural network capable of mimicking fruit fly’s visual system
The simple structure of fruit fly’s visual system is capable of reliably distinguishing between individuals based on sight. A team of researchers from University of Guelph and the University of Toronto, Mississauga developed a neural network that replicates fruit fly’s visual system to distinguish and re-identify flies. Expertise in fruit fly biology and machine learning collectively helped to build a biologically-based algorithm that studies low-resolution videos of fruit flies to determine whether it is physically possible for a system to mimic fruit fly’s visual system to distinguish and re-identify flies.
Fruit flies have small compound eyes and the visual system takes in a limited amount of information. It was previously assumed that once the image is processed, a fruit fly can distinguish very broad features. However, according to new findings, fruit flies can boost their effective resolution with subtle biological methodology. This has led researchers to theorize that vision contributes significantly to the social lives of flies. In the current research, the team suggested that the fruit flies’ visual system is similar to a Deep Convolutional Network (DCN).
The computer program built by the team contains the theoretical inputs and the processing ability as a fruit fly. The system was trained on video of a fly over two days and the team revealed that the system was able to reliably identify the same fly on the third day with an F1 score, which is a measure that combines precision and recall, of 0.75. The project was funded by a Canadian Institute for Advanced Research (CIFAR) and were conducted at the University of Toronto Mississauga lab of Joel Levine, a senior fellow in the CIFAR Child & Brain Development program. According to Levine, the approach of pairing deep learning models with nervous systems can help to understand communications of neurons with each other in animals. The research was published in the journal Public Library of Science (PLOS) One on October 24, 2018.