For centuries honeybees have attracted human fascination. Not only are they polinators and producers of the nutritious sweet substance but also members of an amazing social structure. As social insects, honeybees form large family groups which allows them to perform tasks impossible to solitary insects, such as construction of complex nests, division of labor, or group defense. The details of the bee communication system, regulation of supply and demand and allocation of tasks are unknown.
In this work I leverage the recent advances in deep learning to build image analysis methods for multi-object tracking. I devised a method for finding orientation and location of densely packed beed on a surface of a 2D observation beehive. Current method for finding bee identities allows for tracking hundreds of individuals over a short time span.

Bozek K, Hebert L, Mikheyev AS, Stephens GJ Towards dense object tracking in a 2D honeybee hive Computer Vision and Pattern Recognition 2018

C. elegans

I am also interested in the behavior of C. elegans. Traditionally, behaviors are categorized by observers trained to identify relevant elements. In ethology, study of animal behavior, these would include feeding, fighting, or mating. Here I am exploring alternative to the subjective categorization-based ways to quantitatively represent behavior. Recent comprehensive, category-free ways of quantifying animal behavior have pointed to few general principles: behavior is low-dimensional – it can be described with fewer values than the morphological degrees of freedom; also it is stereotyped, it occurs in distinct, repetitive patterns.
Here I use unsupervised deep learning methods to find representations of C. elegans posture space. Similar to other organisms, behavior of this nematode can be comprehensively captured with only few variables. It is also stereotyped allowing to realistically simulate its movement with recurrent neural network models operating on the low-dimensional worm representations (see the animation above!). Novel, descriptions of C. elegans behavior could potentially better map onto other variables of interest, most prominently on the genetic, neural, or gene expression variation.