计算机科学
杠杆(统计)
浮游动物
人工智能
深度学习
机器学习
转化式学习
限制
生态学
心理学
教育学
机械工程
生物
工程类
作者
Erik Bochinski,Ghassen Bacha,Volker Eiselein,Tim J. W. Walles,Jens C. Nejstgaard,Thomas Sikora
标识
DOI:10.1007/978-3-030-05792-3_1
摘要
Ecological studies of some of the most numerous organisms on the planet, zooplankton, have been limited by manual analysis for more than 100 years. With the development of high-throughput video systems, we argue that this critical bottle-neck can now be solved if paired with deep neural networks (DNN). To leverage their performance, large amounts of training samples are required that until now have been dependent on manually created labels. To minimize the effort of expensive human experts, we employ recent active learning approaches to select only the most informative samples for labelling. Thus training a CNN using a nearly unlimited amount of images while limiting the human labelling effort becomes possible by means of active learning. We show in several experiments that in practice, only a few thousand labels are required to train a CNN and achieve an accuracy-level comparable to manual routine analysis of zooplankton samples. Once trained, this CNN can be used to analyse any amount of image data, presenting the zooplankton community the opportunity to address key research questions on transformative scales, many orders of magnitude, in both time and space, basically only limited by video through-put and compute capacity.
科研通智能强力驱动
Strongly Powered by AbleSci AI