Drosophila melanogaster or fruit flies can reliably distinguish individuals through their sight, despite having a simple visual system. In a recent study, researchers have developed a Artificial Fly Brain and neural network that replicates the visual system of fruit flies to differentiate and re-identify the flies. It also provides an evidence that vision of a fruit fly is clearer than previously predicted.
According to the researchers, the new study will enhance the work of various labs around the world that use these flies as a model organism to understand how they change over time.
In the latest project, research groups from University of Toronto, Mississauga and Guelph University, bring together machine learning and biology of fruit fly to develop a Artificial Fly Brain and biologically-based algorithm. Through low resolution videos of Drosophila, the scientists obtained the algorithm to test if it is physically possible for a system with innumerable limitations to re-identify the individual flies.
The study, published in the recent issue of journal PLOS ONE, explained that Drosophila’s compound eye is very small with certainly low resolution and takes in an estimated 29 units square of visual information. In the traditional view, the level of detail was limited to very broad features such as movement and regular patterns.
However, recent experiments revealed that fruit flies can increase their effective resolution with subtle biologic tricks. This led the scientists to believe that fruit flies’ vision can significantly contribute to their social lives. The new findings in combination with a discovery that the structure of fruit flies’ visual system is similar to a Deep Convolutional Network (DCN) influence the scientists to develop a fly-brain model that can identify individuals.
The new computer program showed same theoretical input and processing ability as a D. melanogaster and was presented with video of fly for two days. On the third day, the fly-brain model could achieve a relatively high F1 score of 0.75, which was slightly lower than the scores of 0.85 and 0.83 for algorithms. To remove the ability to fly-brain model to measure absolute size and shape, the image was resized to 25%. Impressively, the model outperformed humans and achieved a F1 score of 0.55. Further, it could unmistakably identify male and female.
First author Jon Schneider said that the study direct to the tantalizing possibility of fruit flies to be able to identify individuals rather than just recognizing broad features. “It’s exciting to find a problem where algorithms can outperform humans”, said Graham Taylor, a machine-learning specialist.