University of Arkansas, Fayetteville


Adversarial training, domain adaptation, semantic segmentation, semi-supervied learning


It is well known that arthropods are the most diverse and abundant eukaryotic organisms on the planet. Museum and research collections have huge insect accumulations from expeditions conducted over history that contain specimens of both temporal and spatial value, including hundreds of thousands of species. This biodiversity data is inaccessible to the research community, resulting in a vast amount of “dark data”. The primary objective of this study is to develop an artificial intelligence-driven system for specimen identification that greatly minimizes the time and expertise required to identify specimens in atypical environments. Successful development will have profound impacts on both ecology and biodiversity sciences as it will increase the resolution for ecological studies and allow us to work through the backlog of insect collections, unlocking tremendous amounts of biodiversity data. Development of the system will address multiple challenges in deep learning, including problems associated with limited training data and moving from known domains into unknown. The cutting-edge AI solutions will be a final component in a smart specimen identification system scalable in multiple platforms and across geographic region.